Production, Evaluation, and Preservation of Experiences:

Constructive Processing in Remembering and Performance Tasks

Bruce W. A. Whittlesea,

Simon Fraser University

Send correspondence concerning this article to :
Bruce W. A. Whittlesea,
c/o. Dep't of Psychology,
Simon Fraser University,
Burnaby, B. C.
Canada
V5A 1S6

Memory is the faculty of learning, the capacity of mind to be changed by an experience, such that its potential to respond to stimuli is altered. Memory is the sum of all one's experiences: it is the locus of all skills, all knowledge. Memory controls all of behavior except reflex activity. In understanding human behavior, it is therefore extremely important to have a proper understanding of the structure and function of memory. Unfortunately, memory is involved in a bewildering variety of activities, differing on many dimensions: perceptual, conceptual and motor activities, assimilation of information and control of performance, social, intellectual and emotional interactions, simple responses and complex chains of behavior, abstract reasoning and automatic reactions. Each of these activities and dimensions has its own quirks and special conditions. It is thus a considerable challenge to educe simple, fundamental principles of memory as a whole.

In attempting to deal with the complexity of memory's specific interactions with various tasks and stimuli, investigators of mind have long treated memory as serving two basic and qualitatively different functions, namely remembering and the control of performance. Remembering is the use of memory as a record of events to reconstruct past experience, in the act of recall or recognition: It is the attempt to think about some particular experience of an object, in some specific context. Isolated or unusual events are remembered better than regular, normal events: it is hard to remember any specific experience of riding a bus, distinct from all other such experiences, unless something unique happened during a particular journey. This ability to think selectively about particular past events appears to imply that memory retains knowledge of specific events, which can be retrieved to consciousness.

In contrast, other tasks, such as classification and identification, require the person to perform conceptual, categorical or perceptual judgments, without reflecting on the specific past events in which they acquired the knowledge that enables them to perform the judgments. These tasks call on memory to provide information about the meaningful properties of objects, such as their identity or utility to satisfy some current purpose. Performance in these tasks is optimal for objects that are regular, average or typical of their class. For example, people classify a typical fruit, such as an apple, faster than an atypical fruit, such as a fig (Rosch & Mervis, 1975); orthographically regular words are identified more efficiently than orthographically irregular words (Wheeler, 1970). Such observations seem to imply that memory also possesses abstract, general knowledge about concepts and categories of objects.

Thus the fundamental fact about memory seems to be that it possesses two qualitatively different types of knowledge, specific and general, selectively used to perform two types of task, one with respect to particular events and the other with respect to the generic identity of objects. This dichotomy of knowledge types entails a host of other assumptions about the mechanisms by which each is acquired, stored and applied. In fact, remembering and control of performance appear to depend on completely separate memory systems, operating under qualitatively different sets of principles. This reasoning is the cornerstone of theories of memory such as Anderson's (e.g., 1980) associative network account and Tulving's (e.g., 1985, 1995) multiple-systems account.

I contend that dichotomizing memory in this way seriously miscasts its essential nature, and results in a series of mistaken ideas about each of memory's subsidiary functions. Instead, in this chapter I propose that the prime functions of memory are the construction of psychological experiences and the preservation of those experiences. The construction of an experience of a stimulus is performed through the same set of principles both in remembering specific events from the past and in identifying or classifying present stimuli. The detail of the constructive process on different occasions is extremely various, responding to differences in task, stimulus structure, context, and the availability of similar prior experiences, but the fundamental principles are the same in all cases. Further, I contend that memory preserves only one type of information, namely whatever a person does in constructing an experience of a stimulus. Fundamentally, I argue, memory is unitary in structure and operation. I will describe a formulation of memory, which I call SCAPE (for Selective Construction And Preservation of Experiences). This theory is a marriage of ideas from many sources, including instance theory (e.g., Brooks, 1978; Medin & Schaffer, 1978; Jacoby & Brooks, 1984), the episodic-processing account of concept acquisition (e.g., Whittlesea & Dorken, 1993), the attribution theory of remembering (e.g., Jacoby & Dallas, 1981; Jacoby, Kelley & Dywan, 1989; Whittlesea, 1993), skill transfer (e.g., Kolers & Smythe, 1984) and transfer-appropriate processing (e.g., Morris, Bransford & Franks, 1977; Roediger & Challis, 1992; Masson & MacLeod, 1992).

Separate-systems accounts assume that memory contains two qualitatively different kinds of information, respectively of specific events and general knowledge. They thus have to explain how these two types of knowledge are acquired separately, preserved separately and applied selectively in different tasks or situations. Event knowledge is thought to be acquired in the act of processing stimuli in specific contexts for particular purposes. Memory registers those aspects of the stimulus and context that happen to be processed during that encounter, given the purpose of the encounter, the attentional resources available, and the availability of prior experiences to guide current processing. Event knowledge is thus encoded directly from experience: it is a record of what the person actually experienced in the encounter (the encoding variability principle; e.g., Craik & Lockhart, 1972). Remembering consists of activating a specific representation, independent of all other representations that include similar elements: for example, remembering that TABLE was in a training list requires the person to access that event distinct from all other experiences of that word. Accurate retrieval thus depends on the distinctiveness of the original experience, and on the availability of cues reinstating distinctive aspects of that experience (the encoding specificity principle; e.g., Tulving & Thompson, 1973). The acquisition, contents and retrieval of event knowledge are thus sensitive to the idiosyncratic properties of particular experiences of stimuli.

In contrast, performance in classification and identification tasks is sensitive to general, summary properties of stimuli, such as frequency, typicality and regularity. These attributes are abstract properties of events taken as classes, not properties that could be directly apprehended in any actual encounter with an object: They emerge across a succession of events, but are not present in any one of them. Sensitivity to these properties must therefore mean that memory has some mechanism that computes general properties across particular experiences, abstracting the common meaning or structure of successive experiences, and leaving behind idiosyncratic properties, such as context and purpose. This abstraction is unconscious and automatic: It is driven by the structural similarities between members of a class rather than by intention. This general knowledge is preserved separately from knowledge about the events from which it was abstracted, in the form of logogens, prototypes, rules, or linkages in a semantic network. Initially labile, these abstract summaries become strong and stable as they accrete the common properties of hundreds or thousands of instances of a category. Once acquired, they take control of the perceptual and cognitive processing of new stimuli, providing the person with knowledge of the meaning, class or identity of the stimuli. The efficiency and success of this processing depends on the extent to which the properties of a new stimulus match those represented in the abstract summary, giving rise to regularity and typicality effects.

Accounts of memory based on such automatic abstraction of general knowledge have been proposed by many investigators, studying quite different aspects of performance, including dissociations between remembering and identification (Tulving, Schacter & Stark, 1982; Tulving, 1995), implicit learning (Reber, 1989, 1993), concept acquisition (Rosch, 1978), semantic priming (Meyer & Schvaneveldt, 1971), word identification (Paap & Noel, 1991), repetition blindness (Park & Kanwisher, 1994) and many other phenomena. It is still the dominant theory of memory representation, as demonstrated by the fact that it is taught, in all of its ramifications, in most of the current textbooks, with only minor acknowledgment of an alternative.

Selective Construction and Preservation of Experiences: Outline of the Account

Despite the wide success of separate-systems accounts in explaining cognitive phenomena, I argue that those accounts are based on three fundamental errors about the structure and function of memory. The first error is that there is no mechanism in memory to perform chronic, unconscious abstraction.1 In consequence, there is no semantic memory, no type representations, no lexicon, logogens, prototypes or other forms of automatically-computed summary information. Instead, memory simply preserves particular experiences. Each encounter with a stimulus will produce some specific experience, dictated by the physical structure of the stimulus, the processing performed on it to satisfy the current purpose, the physical and meaningful context in which it is encountered, and the availability of prior similar experiences to guide current processing. Memory preserves that experience, and that is all that it records. In consequence, there is no distinction to be made between a semantic and an episodic memory system.

The second error is the idea that memory is a store of knowledge: that is, a library where facts and concepts are stored, which can be accessed as needed to recall prior events or identify current objects. Separate-systems accounts not only make a distinction between general and particular knowledge, but also between declarative and procedural knowledge (e.g., Squire, 1992; Tulving, 1995). "Declarative knowledge" is meant to be the body of conceptual and factual information that one has to draw upon to identify objects or recall events, whereas "procedural knowledge" consists of the skills necessary to perform operations on stimuli, such as riding a bicycle. However, I will argue, in any interaction with a stimulus, representations of prior experiences guide current processing of the stimulus, producing a mental and/or motoric event. Naming a bicycle or classifying it is as a means of transportation is as much a skilled operation as riding it (cf. Kolers, 1973; Kolers & Smythe, 1984). The fact that one operation consists of producing a verbal label for a stimulus and the other of producing a motor response to the stimulus is not a fundamental difference. In each case, the person encounters the stimulus for some purpose and in some context: that stimulus complex cues representations of prior similar experiences with similar stimuli, which in turn guide processing of the current stimulus. I argue that memory does not directly contain any knowledge about WHAT things are. Instead, memory preserves records of OPERATING on stimuli, constructing cognitive events. In turn, those records drive performance in later encounters with other stimuli. In that sense, all knowledge is procedural. The library metaphor had misled us. In fact, memory is dynamic and interactive with the world. It controls performance in interacting with stimuli, rather than existing as a set of passive records to be consulted.

The third error is the idea that remembering consists of the retrieval of a representation of some prior event. In separate-systems accounts, it is this idea that separates memory's function of remembering from its function of controlling perceptual and conceptual performance: remembering consists of finding and elevating a memory representation to consciousness, whereas in performance of the latter tasks, a summary, abstract representation, that is not itself retrieved to consciousness, directs the processing of some stimulus. I argue that nothing is ever retrieved from memory, in the sense of finding a representation and bringing it intact to consciousness. Instead, memory PRODUCES behavior, both overt and mental: and it does so not only in the control of performance tasks like identification, but also in remembering. The mental event that occurs while remembering may be very similar to the actual past event one is trying to remember, if it is selectively guided by the representation of that experience. However, it is not a copy of the representation of that event, stimulated into consciousness by a cue. Instead, it is a construction imposed on the stimulus, a construction guided by the representation of the prior event, but also influenced by expectations and attitudes aroused by the detail of the current stimulus encounter. In general, I argue that all mental events, whether consisting of the identification of a common object, the remembrance of an event, or the experience of pleasure in smelling a rose, are constructions created by the interplay between stimuli, tasks, contexts and prior experiences: No mental event reflects direct knowledge of the stimulus.

I contend that there is one memory system, that preserves all experiences and controls all behavior. Every processing event involves a stimulus compound, consisting of a stimulus (either an external object or an internal mental event), a purpose (either given by some external agency or arising from prior processing), and a context (both the physical, external world and the set of expectations that mind carries to the event). This stimulus compound is the thing on which memory operates. The concatenation of properties of the stimulus compound selectively cues memory representations of events that involved similar tasks, stimulus structures and contexts. The cued traces control the processing of the stimulus compound, producing a new experience. And memory preserves the new processing experience. At the most fundamental level of analysis, this is all that happens in any encounter with the world, whether the purpose of that encounter is to identify or use a stimulus or to remember some prior event.

The processing of a stimulus compound is an essentially constructive event. The stimulus itself possesses a set of physical properties, but not meaning or organization. For example, ORCHID, as a stimulus, is a physical entity. Its physical properties support the computation of many non-physical properties, such as location, extent, unitariness, sound, pleasantness, familiarity, identity and meaning. However, those are not properties of the stimulus object itself: instead, they are constructed from cued memory representations of prior experiences and attributed to the stimulus, in the act of processing the physical stimulus for some purpose.

The construction of an experience of a stimulus occurs through two semi-distinct activities, namely production and evaluation. In terms of the processing performed, the only difference between these activities is whether the target stimulus is an external or internal one: the mechanism of processing is the same in the two cases. However, they are very different in effect. The former permits the person to deal with properties of the external world, and the latter permits them to experience subjective states about their performance.

The production activity consists of attributing properties to a stimulus, properties that are not physically present in the stimulus. Production of these properties begins when the stimulus compound (the concatenation of stimulus structure, purpose, and context) selectively cues representations of similar experiences. These cued representations preserve a record of performing specific perceptual and cognitive operations on specific stimulus structures. They control subsequent processing of the current stimulus, directing the construction of an organized percept of the stimulus, the imposition of conceptual properties such as identity on the percept, and the production of a response. The production activity ends with the occurrence of a mental or behavioral event. These end events can be exceedingly various, such as speaking a name in greeting a friend, the coming-to-mind of the meaning of a word or phrase in reading a novel, the act of patting in encountering a dog, or the perception of relationship in admiring a picture. They can be as simple as imposing a global perceptual organization on a circle of dots or as elaborate as picturing oneself using a stimulus object in an imaginary environment. The mental or behavioral event produced in a stimulus encounter can be an end in itself, as for example in classifying a stimulus as a WUG in a psychological experiment. However, the event will often become part of the context of another stimulus complex, and help to determine the processing of another stimulus. For example, in reading a novel, the coming-to-mind of the meaning of one word in a sentence influences the interpretation of the next word, and so on.

The production of mental events is always controlled by representations of specific prior experiences, selectively cued by the details of the stimulus complex. This is often not obvious in performance in tasks such as naming and classification, which appear to be fairly impervious to changes in the context or perceptual manifestation of test stimuli. Instead, as discussed earlier, performance in such tasks is correlated with the structural similarity of a stimulus to the mass of past experiences of the same class. However, I will argue in Part I of this article that this apparent stability is illusory: it occurs not because production of identity is insensitive to contextual detail or driven by an alternate knowledge base, but because similarity to specific experiences and to abstract summaries of those experiences is usually confounded in the experiments used to study those tasks. When specific experiences diverging from the average are supplied in an experiment, performance in identification and classification is observed to be controlled through an interaction of the specific cues offered at test with the specific characteristics of prior experiences.

The act of remembering is also a constructive activity. In this case, the person is not asked to produce the identity or class of a stimulus, but instead some property that was associated with the stimulus on a prior occasion. In recognition, a person is given an item as a stimulus and asked to produce the identity of a context in which it has been encountered previously: in recall, the person is given a context as a stimulus and asked to produce the identity of an item that occurred earlier in that context. These properties are constructed in the same way that the meaning or class of an item is produced in identification and classification. The specific characteristics of the present stimulus compound selectively cue representations of prior experiences, which direct the construction of a mental event. The only difference is that, owing to the demand of the task, the person constructs a mental experience of some context, rather than a mental experience of the identity of the stimulus. The constructive nature of remembering is discussed in Part II of this article.

The production function permits people to interact efficiently with the world, generating identities and remembrances. When such productions are performed fluently, one can usually trust them to be accurate. If the word "dog" comes to mind on seeing a small, hairy quadruped, it is probably the correct label; if I am trying to remember where my glasses are, and an image of them lying on the hall table comes to mind, I will probably find them there. The accuracy of these productions stems from the selectivity of the stimulus compound in cueing similar experiences, which will usually involve relevant information.

Such fluent productions do not seem to be accompanied by any strong feeling-state: in fact, they are often barely noticed. In reading the word GRASS in a novel, one accepts the coming-to-mind of the meaning of GRASS without question, and without experiencing any particular emotion. One does not even experience a feeling of confidence that one is correct: one simply understands the word, and passes on. In remembering to buy milk on the way home, I may experience annoyance at having to go out of my way, but the remembrance itself is not accompanied by any strong feeling, such as a feeling that the idea of buying milk is awfully familiar. So long as the production occurs fluently, one "just knows" the information produced, and proceeds to the next activity.

However, there are many occasions when the production operations fail in a surprising way. For example, one can think of a familiar meaning, but fail to produce the word that has that meaning, or encounter a familiar face but fail to produce the context from which one knows that person. On these occasions, people experience a tip-of-the-tongue feeling, or a feeling of familiarity, that can be very powerful (and annoying). They also interrupt ongoing activities to concentrate on the offending stimulus: stop speaking to concentrate on that word, or stop looking around the bus to stare at that individual. Such failures of production can occur in a variety of ways. For example, the production function may succeed in generating a complete mental event, but may do so with unexpected fluency or unexpected awkwardness. Alternatively, I may discover that the mental event was in error: at closer range, the quadruped turns out to be a cat, or the glasses are not on the table. An external agency, such as an experimenter, may also question the result of the production operation, asking "Did you produce the right answer? Are you sure?". Under any of these circumstances, memory performs a second function, namely evaluation. 2

The evaluation function is performed through exactly the same mechanism as the production function: a stimulus compound is processed within a context and for a purpose, guided by cued representations. However, the object of attention in this case is the quality of the mental event that was just produced in encountering a stimulus (the fluency, completeness or elaborateness of processing), rather than the external stimulus itself. The purpose is also different: instead of producing a response to an external stimulus, the object now is to evaluate the goodness or source of the mental event that just occurred. For example, in a recognition experiment, a subject shown a test word may experience a mental image of seeing that word in the training context. On that basis, they can infer that the item is old. However, they could produce such a mental event whether or not they had actually experienced the test word in that context, just as one can imagine the word HORSE written in red gothic letters, even if one has never seen it that way. To decide between these alternatives, the subject may evaluate the fluency with which that image comes to mind, or the clarity, elaborateness or completeness of that image. They could expect that an actual prior experience of the word in that context would sponsor more vivid, efficient and detailed production of the mental image than would ad hoc construction. Similarly, in classifying objects, a person can use the ease with which a category label comes to mind as an indicator of whether that label is correct.

Thus people can use the efficiency of the production of a mental event to judge the source or accuracy of that event. This evaluation function enables memory to make attributions about either the nature of the stimulus or about general or specific aspects of their previous experience of the stimulus. The evaluation process can be performed deliberately and consciously, but also occurs automatically and unconsciously when processing is surprisingly fluent or nonfluent.

This evaluation is a heuristic, inferential process. Just as many different meanings or organizations can be imposed on a stimulus in the production function, depending on the task and context, so many different meanings can be imposed on unexpectedly fluent or nonfluent processing. In the context of a recognition task, fluent processing may be attributed to the influence of past episodes, and be experienced as a feeling of familiarity: in the context of a classification judgment, the same fluent processing may instead be attributed to structural goodness of the stimulus, and be experienced as a feeling of rightness. Unexpectedly non-fluent processing may be interpreted either as novelty of a stimulus or as an error in producing a name or behavior.

The effect of evaluative processing is thus to produce feelings about stimuli, feelings of rightness and wrongness, pleasantness and unpleasantness, novelty and familiarity and so on. These feelings provide an important second basis for performing both remembering and non-remembering tasks, beyond the simple production of mental events. They also appear to serve an alerting function, warning that there is something wrong or special about the processing just conducted. They are not just information about the nature of the stimulus (that it is unpleasant or familiar): they are motivating, driving the system to solve a problem. For example, when a face at a bus-stop feels familiar, one can guess that one has seen that person before (the information function): but one will also be driven to search for the reason that face feels familiar, attempting to produce the context from which that face is known, or to identify the friend whose face that face resembles. If one succeeds in producing the source of the feeling (e.g., produces a mental image of seeing that face on the clerk at the corner store), the feeling of familiarity disappears, replaced by "just knowing" the person. If one fails to produce the source, then the feeling of familiarity persists, warning about an unsolved problem. The evaluative function thus serves as an important backstop to production, alerting the system to stimuli that may be of special significance.

To summarize, performance in remembering tasks is sensitive to the episodic detail of particular experiences, whereas performance in non-remembering tasks is often observed to be sensitive to general properties of the mass of prior experience. According to separate-systems accounts, this evidence means that memory serves two functions, retrieving specific events from the past and producing knowledge about present objects. These functions are served by separate memory systems that preserve qualitatively different kinds of information, and acquire and apply their information through qualitatively different principles. In contrast, the SCAPE account argues that the differences in performance in remembering and non-remembering tasks are due to the differences in the cue compound and the interpretive contexts made available in the tasks. It also argues that memory performs two functions, construction and preservation of experiences, but that those functions occur in the same way in remembering and performance tasks. Fundamentally, there is only one kind of representation in memory, and one set of processing principles: a current stimulus compound selectively cues similar prior experiences; those experiences guide current processing; memory preserves a representation of that experience; and the new representation serves as a resource for perception and performance on further occasions.

The principles of the SCAPE account are simple, but have wide and diverse implications. I will illustrate the theory through six themes that have controlled my research program. In Part I, I attempt to demonstrate the constructive nature of performance in tasks such as classification and identification. The experiments show that automatic, unconscious abstraction is an unnecessary and insufficient assumption, and that instead, performance in such tasks depends on the preservation of particular experiences of particular stimuli. They demonstrate the utility of the "episodic-processing" account of acquisition, representation and production, on which SCAPE theory is partly based. In Part II, I attempt to illustrate the constructive nature of remembering. The experiments also demonstrate the interplay between the production and evaluation functions of memory, and the "attribution" account of decision-making which makes up the remainder of SCAPE theory.

Part 1:

Constructive Production

Wittgenstein (1953) pointed out that many natural categories have an internal family resemblance structure, consisting of the fact that each member shares some features with some other members, although any two members may share no features in common. For example, the overlap of features in the items {ABC, CDE, EFG} binds them together as a set, distinct from the set {KLM, MNO, OPQ}. Further, the overlap of features that defines the category means that some features, like C and E, are more common than others, so that items bearing them, like CDE, are more typical of the set as a whole than are other items.

Rosch (e.g. 1977, 1978) demonstrated that this internal structure of categories influences people's behavior. Rosch and Mervis (1975) asked people to rate the typicality of members of the FURNITURE category, and also to list the features of each item. They observed that people's typicality judgments could be predicted from the number of features that any member shared with other members of the category. For example, the item CHAIR, which shares features with many other members, including TABLE, DESK and SOFA, was rated as highly typical of the category, whereas CARPET, which shares few features with any other member, was judged to be atypical. Rosch concluded that people's judgments of typicality must be controlled by the distribution of features across members of the category. However, people never directly encounter that distribution: instead, they only encounter the individual members from which the distribution can be calculated. The question was thus how people come to be sensitive to the abstract properties of the category as a whole.

In any such category, there is usually one item, called the prototype, that shares more features with all other items than any other does, and so is most typical of the set as a whole. Posner and Keele (1968) observed that the prototype has a favored status among category members: even if it is not shown in the training phase of an experiment, it is classified as well as any training item, and better than novel items. Moreover, the probability of correctly classifying other members of the category can be predicted from their similarity to the prototype (e.g., Franks & Bransford, 1971; Neumann, 1974; Rosch, Simpson & Miller, 1976). This correlation of typicality of test stimuli with accuracy of classification convinced many researchers that the prototype must be directly represented in memory, even if it was never presented (e.g., Rosch, 1977, 1978; Homa, Sterling & Trepel, 1981). It was thought to be computed through an unconscious, automatic abstraction mechanism, which tallies the frequencies with which various features occur in successive training stimuli. The prototype would embody the set of most typical features, and become the standard that people use to classify other members of the category.

Despite wide acceptance of the idea that category learning proceeds through the abstraction of prototypes, the evidence for it remained correlational. There was thus the possibility of an alternate explanation of the relationship between typicality and accuracy. Brooks (1978) and Medin and Schaffer (1978) suggested that the correlation between typicality and accuracy could be explained equally well if subjects simply encoded each presented instance of the category, and compared test items to the set of encoded instances. Because the prototype is the average of those instances, a test item that is similar to the prototype is also necessarily similar to many instances of the category, whereas an atypical item would be similar to fewer instances, and also probably less similar to any specific instance. The prototype itself is more similar to other instances than is any other instance, and so would benefit most from this comparison in a test. The question thus became whether the relationship between typicality and accuracy is direct and actual, resulting from abstraction of the prototype during the learning phase, or indirect and potential, resulting from simply encoding the training instances.

No direct evidence of prototype abstraction has yet been presented. However, there is now much support for the alternative hypothesis, that typicality effects in category learning result from encoding and preserving representations of individual training instances. For example, Whittlesea (1987) unconfounded the typicality of test stimuli (their similarity to the prototype) from their similarity to individual training items. I created two categories of items, based on the prototype strings FURIG and NOBAL. Training stimuli all differed by two letters from their prototype; for example, the FURIG-based training set were {FEKIG, FUTEG, PURYG, FYRIP, KURIT}. Subjects were shown each item three times, pronouncing and copying them on each occasion.

Test items were presented tachistoscopically: the dependent variable was the subjects' accuracy in identifying the component letters. In this test, there were two types of novel stimuli. One type (e.g., FUKIP) shared three letters with its prototype, the other (e.g., PEKIG) shared only two. However, the first type shared only three letters with the most similar training item (FEKIG), whereas the second shared four. This provided a critical test of the alternatives: If people abstract prototypes across instances, and use them to perform judgments about category members, then performance should be superior for the more typical items, whereas if they simply preserve representations of the individual stimuli, then performance should be superior for items more similar to actual training instances. Contrary to prototype theory, but in support of instance theory, the latter pattern of performance was observed.

Further experiments in this series demonstrated that, although the most similar training stimulus exercises most control over performance on a test item, other instances can as well. For example, in one study, subjects were shown two types of novel test items (e.g. FUKIG and PEKIG) that were equally similar to the most similar training instance (FEKIG). However, FUKIG was more similar to the remaining instances of that category than PEKIG: it was also more accurately identified. In five further experiments, I observed that the subjects' relative accuracy of identification could be accurately predicted from the simultaneous similarity of those items to the set of training instances.

The observation that most impressed early investigators of the typicality effect was that the prototype item could be classified with equal, or sometimes greater, accuracy than items that had actually been seen in training. In contrast, in the two studies just reported, training items (e.g., FEKIG) were also presented in test, and were always better identified than novel test items, even items closer to the prototype (e.g. FUKIG). However, in those studies, training items presented again as test items had the distinct advantage of being identical to one out of the ten items presented in training: In consequence, their average similarity to the set of training items was in fact greater than that of more typical items. I reduced this advantage by increasing the stock of training items to fifteen per category. Now, any training item that occurred again as a test item was identical to only one out 30 training items, and instead items closer to the prototype were more similar on average to the training set. In this case, the pattern reversed: the novel, more typical items were at last classified more accurately than items the subject had actually seen in training.

The results of these studies lead to several important conclusions. First, the idea that categories are represented by an abstracted prototype is unnecessary and insufficient to explain people's performance in tasks involving members of well-structured categories. Instead, memory preserves representations of particular instances of categories, and those representations control performance on novel instances in a perception or classification test. Second, neither similarity to a prototype nor prior presentation as a training instance is sufficient to predict relative success in identifying test instances. Instead, judgments about category members are driven simultaneously by all of the instances that memory has previously encoded. In small domains, or domains containing relatively distinctive members, this can cause instances that were actually experienced previously to be processed more accurately or efficiently than more typical instances. In contrast, in larger domains, containing less distinctive members, items that are more typical of the category as a whole will be more similar to the set of encoded instances, and will thus be processed optimally.

Finally, these results stand as a warning against interpreting correlations of performance with general properties of a domain as evidence that people have actually abstracted those properties. Typicality is ordinarily a good predictor of performance in family-resemblance categories. However, no matter how strong the correlation, it is not direct evidence about the actual basis of performance. Unfortunately, as described below, investigators of numerous phenomena of category and concept learning have accepted such correlational evidence as evidence of automatic, unconscious abstraction, without performing critical tests. To understand the structure and operation of memory, it is necessary to perform direct experimental manipulations of the conditions of learning and test.

The original motivation of instance theory, as proposed by Brooks (1978) and Medin and Schaffer (1978), was to account for people's ability to generalize from experience of a set of instances to novel instances of the same category: That is, it was a theory of concept formation and classification. Originally, the theory consisted of the idea that people learned about categories by learning their instances, as illustrated in the last section. That idea turned out to be very powerful, and has proved to be applicable in all manner of tasks and domains. Over the years, that original conception has been expanded and modified as its application was extended from learning about family-resemblance categories to repetition priming, word identification, implicit learning, and also to remembering. In particular, instance-minded investigators began to realize the importance of context, rather than of instances treated in isolation, and of the specific processes applied to stimuli in encountering them for different purposes. The focus shifted from thinking primarily about the structure of stimuli and the structural similarity of category members to thinking about the experience of a stimulus in some task and context, and the similarity of that experience to other experiences in the same or some other task and context. I chose to refer to this idea as episodic-processing theory (e.g., Whittlesea & Dorken, 1993) rather than instance theory, to reflect the new emphasis.

In fact, a similar idea emerged in parallel in a different literature. Morris, Bransford and Franks (1977) proposed transfer-appropriate processing theory, which states that performance in memory tests succeeds to the extent that the mental operations required in the test match those the subject performed on that item earlier. As originally conceived, transfer-appropriate processing was a theory of remembering, of recognizing the same item on a later occasion. More recently, it has also been applied to cases in which the specific nature of a prior experience influences later non-remembering performance on that item, such as identification (e.g., Roediger & Challis, 1992). From their different roots, episodic-processing and transfer-appropriate processing theory are converging.

At the time that Brooks (1978), Medin and Schaffer (1978) and Morris, Bransford and Franks (1977) proposed their respective theories, concept formation and remembering were treated as utterly different topics: they involved very different tasks, and respectively focused on learning for the future versus retrieving the past, and the ability to handle generalization between stimuli versus knowing the contexts of specific events. Concept formation studies tended to focus on the structure of categories, and the structural relations among instances; remembering studies focused on processing variations at training (e.g. levels of processing; Craik & Lockhart, 1972) and test (encoding specificity; Tulving & Thompson, 1973). Investigators of one topic rarely thought about the other. However, during the 1980s, numerous investigators began to break down the barriers between these areas, and began using tasks that had been previously considered more appropriate for the other area. For example, Jacoby and Dallas (1981) trained subjects under a levels-of-processing procedure, but then tested them not only in a recognition task but also in a flash identification task. Unlike recognition, the identification task does not require the subject to think about their previous experience of the test items in any way. Tachistoscopic identification had long been used in another tradition, namely word identification, in which it was considered to be a direct measure of a subject's knowledge of the general properties of words (e.g., Wheeler, 1970). Jacoby and Dallas instead used it as an indirect measure of the subjects' specific learning about a word in the training phase of a remembering study, contrasted with the direct measure of asking the subjects whether they recognized the item.

The use of indirect measures in remembering studies became increasingly common during that decade. Investigators of remembering began to use fragment completion, listening through noise, pleasantness judgments and a host of other non-remembering tasks to find out about the learning that permits people to recognize specific events. At the same time, investigators of concept formation began to use a variety of indirect measures of category learning, including pleasantness judgments and recognition, in addition to the traditional classification task: these measures are indirect in the sense that they ask the subject to evaluate category members as individual entities or events, rather than to classify them as examples of a category.

Very broadly (and subject to important exceptions), performance on these indirect measures was found to be different in some important ways from that on direct measures, but similar in other important ways. For example, Jacoby & Dallas (1981) observed that prior experience of an item in a training phase influenced performance in both identification and recognition tests, but that the levels-of-processing training affected only recognition. This kind of dissociation sparked an explosion of investigation and theory. Some investigators, such as Tulving, Schacter and Stark (1982), took it as strong support for multiple separate systems of memory, arguing that the different tasks tapped into qualitatively different knowledge bases.3 Others, such as Jacoby (1983; Jacoby & Witherspoon, 1982), took it as support for a unitary memory preserving highly specific experiences of stimuli, experiences which could satisfy the requirements of different tasks to differing degrees.

Despite this cross-borrowing of tasks, research on how information about isolated stimuli is acquired, stored and accessed has continued largely in isolation from research about categories of stimuli, and principles important in one area are largely ignored in the other. For example, there exist two literatures, one on "implicit memory" (e.g., Roediger & Challis, 1992) and one on "implicit learning" (e.g., Reber, 1993), which have almost no cross-talk despite obvious common interest. The former is characterized by a concern with the encoding and representation of individual stimuli (usually words), and with dissociations between direct (recognition) and indirect (e.g., identification) tests of a specific prior experience of those items. The chief tool used to understand these issues is variations in the processing required to be performed on an item at training and at test. Examining transfer from an experience with one stimulus to performance on a structurally similar stimulus is rare, although transfer along semantic dimensions is commonly tested. In contrast, investigators of "implicit learning" are concerned with how people encode and access information about classes of structurally-related stimuli. The stimuli used are letter-strings, numeral-strings or sequences of tones or colors: words are almost never used, so that semantic relationships cannot be examined. The most common manipulations in that area are of the deep structural similarity between test and training items and the identity of surface features used in training and test. Manipulations of the processing that subjects do in training and test through variations in task demands are rare. Even manipulation of the structure of the training set is rare: for example, most investigations of artificial grammar learning have used one artificial grammar, the one given by Reber and Allen (1978).

In my own studies of "implicit learning", described below, I have used designs that are rare in that paradigm, but common in the "implicit memory" literature. These include manipulating the processing at training and test so that they match or mismatch (as in transfer-appropriate processing studies), looking for dissociations between dependent variables, and comparing direct versus indirect measures of learning. These studies have convinced me that the emphasis on structural relations, so endemic to studies of "implicit learning", has led the field astray.

Implicit learning has been extensively studied using the "artificial grammar paradigm". In this paradigm, stimuli are generated from a grammar (a set of structural rules dictating the sequence and repetition of stimulus elements), generating items such as MTTV, VXTVT and VXVRXRM. The basic "implicit learning" phenomenon consists of the observation that subjects exposed to members of a rule-governed domain can later discriminate above chance between novel legal items and items violating the rules, even though they were not aware of the existence of rules during the training, and cannot state the rules in test (e.g., Reber & Allen, 1978).

Two rival factions have emerged to explain this finding. One, the abstraction camp, argues that memory is directly sensitive to the structure of a stimulus domain: it automatically abstracts general information about the deep structure of the domain across the particular instances of the category that the subject experiences in the training phase. That is, in addition to people consciously learning something about the surface structures of individual items, they also compute some aspects of the general deep structure of the domain, either the rules themselves (Reber & Allen, 1978), or common letter-sequences (Servan-Schreiber & Anderson, 1990; Mathews et al., 1989), or the covariance among features (Cleeremans, 1993).

The opposing faction denies the necessity of abstracting general structure of the domain, either consciously or unconsciously, to produce the observed sensitivity to the rules. For example, Brooks (1978, 1987; Vokey & Brooks, 1992) argued that if subjects simply memorized the training stimuli, as they were instructed to do, they would have some success in discriminating legal from illegal test stimuli, because of necessity novel stimuli that satisfy the rules will be more similar to training stimuli than illegal items. Similarly, Dulany, Carlson and Dewey (1984) argued that the subjects' knowledge consisted of salient subunits of items, segments that captured the effect of the rules, and that the subjects were quite aware of having this knowledge. Perruchet and Pacteau (1991) argued that even encoding no more than some random bigrams from the stimuli would enable subjects to discriminate legal from illegal items, because illegal items often contain pairwise violations, and legal items do not. All of these forms of information are available to be encoded directly in separate stimulus encounters, without any need for computation of abstract properties across items.

Most of the debate over the problem of "implicit learning" has thus concentrated on the structure of the stimuli as individuals and as a group, and on what units of that structure people learn when they do not know the domain has a deep structure. The "instance theory" that I described in the last section would argue that people simply encode the structures of particular exemplars of the domain, and classify structurally similar test instances as legal. In contrast, its descendent, the "episodic-processing account" (Whittlesea & Dorken, 1993), argues that people encode processing experiences, not knowledge structures. According to that account, every learning event consists of processing the structure of some stimulus in some context and for some purpose: different contexts and purposes cause the person to impose different organizations on that stimulus structure. The subject's later ability to perform a classification test depends on the demands of that test, and on how they processed training items, given the task at that time, not just on the structure of the individual instances (cf. Vokey & Brooks, 1992).

To demonstrate this, Whittlesea and Dorken (1993) created two grammars, and required subjects to spell instances of one grammar and to pronounce instances of the other. After this training, we explained that the items were taken from two categories, created by two sets of rules. We then showed subjects novel instances of each grammar, asking them to classify these items as belonging to the category they had spelled earlier, or to the category they had pronounced earlier. Before making that decision, they were asked to pronounce or to spell the test item. This activity was crossed with what they had done earlier with similar items of the same category: We required them to spell half of the items taken from the spelling category and to pronounce the rest, and to pronounce half of the items from the pronounced category and to spell the rest. That is, the processing that subjects performed on test items just before classifying them either matched or mismatched the processing that they had done on members of the same category; but the structural similarity of a test instance to other members of its category was the same however they were processed at training or test. If learning about the structure of training items, either at an abstract, deep structural level or at the level of individual instances or bigrams, is the source of "implicit learning", then our manipulation should have no effect. But if the implicit development of sensitivity to categories is mediated by representations of particular experiences, then we could expect to observe processing-specific transfer.

The subjects had some ability to discriminate the test items of the two categories: Overall, their accuracy was 64%, against a chance rate of 50%. However, their success in classifying the test items depended on whether the task they performed on the test items (spelling or pronouncing) was the same as the task they had earlier performed on other members of that grammar. When the processing at test matched that at training, their success in classification was 67%, but when the tasks were different, their success was only 61%.

In a second experiment, the test task was changed, from discriminating between the categories of grammatical items to discriminating grammatical items (of either category) from nongrammatical items. Again, subjects spelled or pronounced novel test items before classifying them, and the test task was congruent with the training task for half of the grammatical test items. The subjects showed excellent ability to discriminate legal from illegal test items: overall, they claimed grammatical items to be legal on 57% of trials, and nongrammatical items on only 27%. However, once again we observed that the match between training task and test processing task was important. The subjects correctly classified novel items as grammatical on 66% of trials when the training and test tasks matched, and on only 48% of trials when they did not.

These results demonstrate that subjects exposed to well-structured stimuli in an implicit learning experiment do not simply acquire knowledge units at some level of abstraction, such as bigrams, instances or abstract rules. Instead, in satisfying the purpose of the encounter, they impose organization on the structure of the stimuli, an organization specific to the task. Performance on subsequent stimuli does not depend simply on the structural similarity of those stimuli to training stimuli, either individually or as set, but instead on the similarity of those structures as processed in the training and test tasks. Successful transfer to novel stimuli depends on having performed appropriate processing in the training.

Proponents of the abstraction hypothesis agree that such results do demonstrate that some part of the transfer observed in implicit learning experiments is due to encoding specific experiences, but argue that there is always residual variance in such experiments that cannot be explained in that way (e.g., Reber, 1993). For example, in the experiments just described, the match-mismatch manipulation does not account for all of the subjects' ability to discriminate legal from illegal items. The residual variance is argued to be due to abstraction of general structure. That argument is weak: the residual variance could as easily be explained through variability in experiencing and encoding stimuli that was controlled by factors other than the experimental manipulation, such as attention, salience, pronounceability and the orthographic similarity of various items to natural words. In any case, it is negative evidence at best.

Whittlesea and Dorken (1993) examined the "mixture of knowledge" idea in a different way. If abstraction of general structure inevitably occurs when people process members of a category attentively , then changes in the purpose for encountering the training items should make no difference to that activity. We asked three groups of subjects to study instances of a grammar, such as RXXCTXT and MTTMRCT, for one of three purposes (none of the groups was informed about the existence of categorical rules). One group was asked to memorize the stimuli for a later test (the usual induction task in implicit learning experiments) by rehearsing them aloud. Another group was also asked to rehearse the items aloud, but told that rehearsal was incidental to their overt task: Subjects were given a 3-digit number to remember, then shown a grammatical instance and asked to rehearse it for ten seconds as a distractor, and then asked recall the number. A third group was asked to say whether each letter of each stimulus was repeated elsewhere in the item. Subjects in all three groups thus processed the entire structure of each stimulus, but understood their task differently.

At test, all subjects were told about the grammar and asked to classify novel items as legal or illegal. All groups were tested on novel legal and illegal items presented in the same letter set as training items, and on the same items presented in a novel set of letters (e.g., the item RXXCRCT being replaced with the item FPPLFLS, which has the same deep structure but a novel surface structure). These tests present different demands. The former can be accomplished simply by encoding some information about the surface structures of some stimuli, coded in terms of the identities of the letters; but the latter requires that the subject have already encoded, or be able to generate at test, some information about the abstract pattern of repetition of letters within items.

All three groups performed above chance in the same-letter test, and to about the same degree. However, in the novel-letter test, the rehearsal-as-distraction group showed no reliable ability to discriminate the items. The memorization group showed a reliable ability to do so, but less than in the same-letter test. The analytic group performed at identical levels in the two tests, and that group performed reliably better in the changed-set test than did the memorization group. That is, the first group showed no ability to discriminate items on the basis of their deep structure: the memorization group could do so to some degree, but less than subjects who had analyzed the patterns of training items. We concluded that the different purposes that subjects had been given for encountering the stimuli had caused them to process training items in different ways, that differentially prepared them for the different demands of the tests. We further concluded that the implicit development of the ability to discriminate legal from illegal items does not demonstrate the operation of automatic, chronic abstraction of deep structure. Instead, that ability, when it occurs, is an incidental by-product of processing training stimuli for a specific purpose.

Some proponents of the abstraction hypothesis argue that memory does not abstract information directly from the stimulus, but instead from the subject's experience of the stimulus (e.g., Mathews & Roussel, 1993). This might go some way to explaining why the subjects who rehearsed instances incidentally in a number-remembering procedure were later unable to discriminate items on their deep structure, even though they had repeatedly rehearsed those items. However, this argument makes the abstraction function seem a bit odd: instead of granting implicit sensitivity to the real structure of the world, as envisioned by Reber and Allen (1978), it instead is a running summary of the average properties of what one has done with stimuli. By this argument, memory codes its experiences twice: once in full episodic detail, and again, computing the average of those experiences. Such redundancy is not impossible, of course: but given that that summary is implicitly represented in the set of episodic representations, so that memory can be sensitive to those average properties without directly computing them (as demonstrated in the last section, on learning of family resemblance categories), it seems peculiar that memory would have developed a special mechanism to perform that activity.

However, such speculations are less important than direct evidence. Whittlesea and Wright (1997) tested the abstraction hypothesis in a different way. We used only one training procedure. All subjects were asked to memorize a set of grammatical instances for a later test (without, of course, being informed about rules). In consequence, all subjects had the same purpose for processing, and would not experience the stimuli in qualitatively different ways. The stimuli were letter strings, such as RMRMCTX and RMCCTXT. We examined what subjects had learned using direct and indirect tests of grammatical sensitivity. The direct test was classification. The indirect test was a pleasantness judgment, in which subjects were asked to describe each stimulus as pleasant or dull.

Pleasantness judgments have been observed to be affected by prior experience of grammatical instances in the same way that classification is. For example, Gordon and Holyoak (1983) found that subjects claimed to find novel instances of a grammar more pleasant than novel illegal items. However, in that case, test items were presented in the same surface features as training items: that is, using the same set of letters in both cases.

Our test items were either identical in deep structure to training items (legal), or simply different (illegal). But in either case, test items were presented in a novel surface structure, being shown either in a novel set of letters or as patterns of color patches instead of letter strings. For example, an item shown as RMRMCTX in training was now shown either as QFQFSPL or as a row of squares colored RED, YELLOW, RED, YELLOW, GREEN, BLUE, BROWN. When test items were presented in novel letters, subjects had a small but reliable tendency to find grammatical items pleasant (about 53%), but a greater ability to classify them as grammatical (about 60%). When test items were presented as strings of color patches, they were still able to classify stimuli as legal or illegal above chance (about 59%), but their pleasantness ratings fell to chance levels (49%). That is, the actual status of the item as conforming to the rules of the grammar influenced the subjects' behavior in the direct, classification test, but not in the indirect, pleasantness test.

These results are difficult to explain through the abstraction account. That account argues that people abstract the common structure of their experiences of a set of stimuli, and apply that knowledge in dealing with novel members of the domain. Although the acquisition of that knowledge may be conditional, based on the subject's specific processing and experience of the training stimuli, the application must be automatic, driven by the structure of the test stimuli. After all, in the standard implicit learning experiment, subjects acquire their knowledge in a memorization task, but are supposed to apply it unconsciously to novel stimuli in a completely different task, namely classification. If the application of that knowledge truly is implicit, then it must also be automatic: in fact, the idea that implicitly-acquired knowledge of general structure could control people's behavior without their awareness was the promised magic of "implicit learning". Moreover, Gordon and Holyoak (1983) showed that memorization of grammatical instances can control performance in indirect tests such as pleasantness judgments as well as in direct tests, if test stimuli bear the same surface structure as training stimuli. If the basis of performance in the classification tests were the automatic application of abstracted knowledge in the presence of familiar deep structure, then we ought to observe an effect on the indirect test whenever we see an effect in the direct test.

Instead, I argue that test performance is based on the cues presented by various test tasks and stimuli, and the ability of those cues to access specific representations of prior experiences: that is, the stimulus compound, not just the stimulus. The classification task does not simply offer subjects legal and illegal stimulus structures: It also comes with the instruction that the discrimination is to be made relative to a specific set of prior stimuli. This instruction cues the subjects to recall any possible information about those stimuli. If the subjects can recall even a few of those items, they can succeed in the classification task above chance, even though the test stimuli are presented in a novel surface structure, by performing what Brooks and Vokey (1991) called an "abstract analogy". That is, having recalled a training stimulus, they can analyze the repetition pattern of that item, and use that information to evaluate the legality of the test stimuli. They thus achieve some abstract knowledge, but about individual items rather than the general properties of the set. Moreover, they do so deliberately and consciously, and at the time of test, rather than unconsciously and automatically during exposure to training stimuli.

In contrast, the indirect, pleasantness judgment does not provide that direct cue back to the training set. Instead, experiences of training items will be cued to control processing in this task only if the stimuli are themselves specifically similar to those experienced in training, not only in formal structure but also in perceptual manifestation. In Gordon and Holyoak's (1983) study, presenting test stimuli in the same letters as training items made them specifically similar in both ways. In Whittlesea and Wright's (1997) changed-letter test, the perceptual manifestation was different, but at least the surface structure still consisted of letters, cueing the previous experiences to a slight extent. However, when we changed to color-patch presentations, the only cue to access the prior representations was the fact that legal test stimuli were of identical deep structure to training items: and that was not sufficient to cue those representations.

I conclude that the phenomenon of "implicit learning" does not require that memory have the function of automatic, unconscious abstraction of structure. Instead, that phenomenon is a by-product of encoding stimuli for particular purposes, experiences which can facilitate other performance in unanticipated tasks on a later occasion. On every occasion of processing a stimulus, one becomes aware of some aspects of processing (those related to the purpose of the encounter), but also performs and records other activities without awareness. The preserved record of those activities can support processing on a later occasion, without the person becoming aware of the source of influence. For example, in reading the word TIGER, I may become aware of the meaning of the word, because that is the purpose of the encounter, but I'm also acquiring a representation of its orthographic code, as it appears in the current font and context, without being aware that I'm doing so. That representation can prime later identification of that word in a similar context: That is, I have implicitly acquired the potential to perform better on a later, unanticipated test of identification. When that priming occurs, I may become aware that I have performed surprisingly well, but not be able to determine the source of that effect (either because I cannot recall the earlier experience, or because I do not understand the effect that that experience could have on my current performance). However, the lack of awareness of the creation and realization of the potential to perform does not call for a separate form of memory. Instead, as I will argue later, it simply demonstrates the general problem that people have in becoming aware of the sources of their performance, a problem that occurs in classification and identification as well as in the act of remembering.

The last two sections have concentrated on experiences of individual stimuli as the basis of performance in nonremembering tasks, to counter the abstractionist claim that performance in those tasks depends on abstracted general knowledge. However, the SCAPE account does not argue that people only possess knowledge about individual stimuli: quite obviously, people can abstract general properties of their experience, and do know information about classes per se. However, the account claims that they do so only in the service of some particular purpose for processing, and not as an automatic, chronic function. Acquisition of such knowledge may be deliberate, as in analyzing instances of a category to discover what they have in common, or computed incidental to performing some other judgment that requires direct knowledge about abstract properties of the domain, such as deciding whether a person is relatively short for an athlete. The means of acquiring representations of such knowledge is different than that of learning about individual stimuli, since it necessarily involves comparison of stimuli and analysis of their common or different properties. The contents of the representations are also quite different, respectively of experiences of individual instances and of class properties of a domain. However, representations of general knowledge are acquired through the same general principles as knowledge about individual stimuli, through the construction of an experience of a stimulus in some task and context, guided by similar experiences in the past, and are accessed in the same general way, through the occurrence of a further stimulus complex that cues similar prior experiences.

This account conflicts with the separate-systems explanation of the acquisition, representation and application of general and particular knowledge. By that account, as discussed earlier, those kinds of knowledge are acquired by different mechanisms and preserved in qualitatively different stores. Moreover, in that account, the chief factor that dictates which type of knowledge is accessed on any occasion is the type of task. Tasks that specify particular contexts, such as remembering, will selectively invoke representations of specific experiences; in contrast, tasks that present stimuli as instances of a general class will cue the application of general knowledge.

Whittlesea, Brooks and Westcott (1994) examined the selective use of general and particular knowledge using the family-resemblance domain described earlier, based on the FURIG and NOBAL prototypes. In this study, the FURIG stimuli were presented as artificial nouns, bearing an -ISM suffix (e.g. FEKIGISM, KURITISM), and members of the NOBAL category were presented as artificial verbs, bearing an -ING suffix (e.g. NEKALING, KOBATING). In all experiments of this series, the subjects were required to learn about the stimuli as individual entities, but were also required to compute abstract knowledge about the set. In one training phase, they pronounced and copied each stimulus (complete with suffix), thereby experiencing it as a unique entity. In another, they were shown the same stimuli, told whether each was a noun or verb, but shown only the stem (e.g. FEKIG). On each trial, they were asked to judge how frequently each letter of the stem had occurred in stimuli of that class on preceding trials, thus computing a running summary of the typicality of each letter across the set.

The question was which form of knowledge, about specific whole items or the typicality of their features, the subjects would apply when they had encoded both, and could use either to perform a test task. This was evaluated through comparison of two types of items, one (e.g. PEKIG) that was more similar to a particular training instance (FEKIG) but less typical of the whole set, and another (e.g. FUKIP) less similar to any training instance but more typical of the category (with prototype FURIG).

At test, one group of subjects was shown whole items (e.g. PEKIGING) and asked whether they were completed with the correct suffix. Another group was shown only the stems of items (e.g., PEKIG), and asked whether they were nouns or verbs. Although these tasks are logically identical, and can be performed using either knowledge about individual features or knowledge of whole items, the former presents itself as item verification, the latter as verification of membership in a class. In the former task, subjects relied on their knowledge about particular exemplars; but when asked to classify the stem as a noun or verb, they instead relied on the typicality of the features. That is, the subjects used particular and general knowledge selectively, depending on whether the task involved treating a stimulus as an entity in its own right or as an instance of a class. This result is consistent with the assumptions of the separate-systems account.

However, in four other experiments, we observed that the use of those forms of knowledge was also controlled by a variety of other characteristics of the task. In one study, subjects were offered a forced choice between two novel stimuli, complete with suffixes (e.g. FEKIGING and NEKALING): they had to decide which had the correct suffix. The stimuli were shown side by side for half of the subjects; for the other half, the stimuli were presented one above the other. Although both tasks treat the items as individual entities, the subjects relied on similarity to particular training items in the former case, but on the typicality of separate features in the latter. In another study, we presented a category name (e.g. VERB) and a whole item (e.g. NEKALING), asking subjects to verify that the stem belonged to that category. Half the subjects received the category name in advance of the item; half received the item before the category name. Although the demand was to classify in both cases, subjects relied on typicality when the category name was supplied first, but on item knowledge when the item was presented first.

Contrary to the separate-systems account, these studies demonstrate that the selective application of general and particular knowledge is not unilaterally controlled either by the structures of stimuli or by the type of task. Instead, selective use of the two levels of knowledge was cued by the specific perceptual organization and cognitive operations that the subject imposed on stimuli in the test. When stimuli were presented side by side, it was easier to compare them by coding each as a whole item, whereas vertical presentation made it easier to compare them a letter at a time. Those organizations of the stimuli were respectively more similar to representations of whole training items or of separate features, and selectively cued those representations, regardless of the demand to verify items as individuals. Similarly, when the name of the class was presented first, it cued class-level information, and caused classification to be based on that knowledge; whereas presenting the item first cued representations of similar items, causing classification to rely on that basis, regardless of the overt type of task.

In general, the selective use of prior experience is controlled by the similarity between a stimulus compound (the structure of a present stimulus, as processed under the perceptual and cognitive operations performed to satisfy the current task) and specific representations of prior experiences (the structures of previous stimuli, as encoded given the operations that were performed at that time). That same principle applies whether the current stimulus is treated as a unique entity, occurring in some specific time and place, as in a remembering task, or as an instance of a generic class, in a classification or identification task. The fundamental organization of memory is not a dichotomy of knowledge types, acquired, stored and applied by different principles. Instead, memory is organized through the specific similarity of the current experience of a stimulus to prior experiences of other stimuli.

Previous sections of this article have dealt with the acquisition and use of information about novel domains. A subject's total experience of such domains during an experiment may amount to no more than thirty or forty instances. However, people deal every day with familiar domains, such as words, that have thousands of separate identities, and uncountable numbers of specific instances. It is worth asking whether processing in artificial domains is representative of processing in familiar, natural domains, such as the identification of natural words and objects.

People's success in identifying a word is to a large extent predictable from its typicality or regularity. For example, orthographically regular words (Wheeler, 1970) and phonologically regular words (Glushko, 1979) are easier to identify than other words: so are words that occur with high frequency (McClelland & Rumelhart, 1981) or belong to high-density orthographic neighborhoods (Andrews, 1992). Such observations have led many researchers to conclude that these abstract properties are directly represented in memory, and directly control the act of identification.

This assumption is the basis of the "dual-route" hypothesis of word identification (e.g., Paap & Noel, 1991). Familiar words have a representation in a "lexicon", a mental table connecting the perceptual manifestations of particular words with their pronunciation and meaning codes. These lexical representations correspond to prototypes in the category literature: They are representations of the essential, abstract structure of the word, summarized across thousands of particular experiences of that word. In consequence, a word like TABLE is represented only in terms of the compound of its features {T, A, B, L, E}; no information about the variations in letter size, font, handwriting style or text context in which the word has been encountered is included in the representation. This form of knowledge is thought to be responsible for the word frequency effect. Unfamiliar words are thought to be identified through the other route, consisting of a set of spelling-to-sound rules that can be used to construct a pronunciation for a letter compound not found in the lexicon. The evidence that people have such rules for constructing pronunciations, even though they cannot describe them, is the same as the evidence for the abstraction of rules from instances of artificial grammars: the subjects' performance is predictable from the rules.

The dual-route hypothesis suffers from the same problem as the prototype- and rule-abstraction hypotheses of category formation: It is based on correlational evidence. The fact that people behave the way they would if they had abstracted logogens or rules does not prove they have done so: The abstract properties of words are confounded with the specific experiences which they summarize. High-frequency words have actually been experienced in more events and contexts than less frequent words; regular words are structurally similar to more words, encountered in more specific contexts, than irregular words. In consequence, effects associated with frequency, typicality, regularity and neighborhood size would occur whether people actually abstract general, deep structural properties across their experiences, or only encode and preserve their specific experiences with words.

These alternatives can be disambiguated by providing subjects with new, distinctive experiences of familiar words. The SCAPE account assumes that each occurrence of a word will be encoded as it was experienced. That means that aspects of the experience that are definitionally irrelevant to naming it, such as the context or font, will be incorporated into the representation, and will influence later interactions. In contrast, the dual-route account argues that the effective knowledge controlling identification consists of abstract information about the structural relationships among features, coded in a lexical entry or spelling-to-sound rule. In consequence, if the dual-route account is correct, then performance in identification tasks should be broadly stable. A single specific experience of a word that has been seen many times in the past should not greatly affect identification of that word on a later occasion. In contrast, if performance depends on the similarity of the current stimulus to prior processing episodes, then identification performance should be highly modifiable, simply by providing one new but distinctive experience.

Whittlesea and Brooks (1988, Exp. 7) investigated the influence of context on word identification. In a study phase, we presented common words either in isolation (e.g. ROUGH, STEEL) or in a simple phrase (e.g. DIRTY BLACK WALL, BIG WOODEN BOX). The question was whether this presentation of a common word, in a specific context, would have an effect on the subjects' ability to identify that word later. In the subsequent test phase, we presented training words in a variety of contexts, some presented alone and some in phrases. These test presentations endured for only 30 ms and were followed by a pattern mask. We observed that a word that had originally been presented alone (ROUGH) was identified very accurately (70% of trials) if it was again presented in isolation at test, but less often if it was now presented in a novel context (DRY ROUGH PLANKS: 34%) or in a context that had been associated with a different item in training (DIRTY ROUGH WALL: 28%). Similarly, words that had been shown originally in a phrase (BIG WOODEN BOX) were well identified when presented in that same context (61%), but less often when presented alone (52%). Subjects had even more difficulty when such words were presented in a new context (NICE WOODEN CHAIR: 43%) and had most difficulty when they were shown in an old but re-paired context (DIRTY WOODEN WALL: 31%).

These results demonstrate the phenomenon of repetition priming (cf. Jacoby & Dallas, 1981). However, they go beyond the standard demonstration of that effect, showing that the effect of a prior experience is not simply to facilitate current performance on a word. Instead, the amount of facilitation of a repeated word depends on the similarity of the test context to the original training experience. Prior presentation of a word in a phrase can even interfere with later identification of that word: target words inserted into contexts originally associated with a different word were harder to identify than words presented in novel contexts or in isolation. We concluded that memory preserves specific experiences of words in particular contexts, and that word identification is critically influenced by these particular prior processing experiences, not just by the average, abstract structural properties of the mass of prior experience.

Forster & Davis (1984) argued that repetition priming effects occur because specific experiences of words are stored in an episodic memory system. In contrast, they argued, identification of words not re-presented in specific contexts is performed through access to the lexicon, which is part of a semantic memory system. Similar arguments, implicating multiple memory systems, have been made by other investigators, including Tulving (e.g., 1985, 1995). By this argument, priming is a special case, that does not inform us about the ordinary activity of identifying a word. However, that argument is based on the idea that priming only occurs under exact re-presentation of a word, which would be a rare and exceptional event. The evidence above demonstrates that prior specific experiences can influence current identification in various ways, even when the test presentation is not identical to the training. That suggests that the phenomenon of priming is much more common than initially thought. In fact, it suggests that every occasion of processing a word is influenced by prior experiences of that word in similar but different contexts, and probably by many prior specific experiences in concert. This effect can only be documented when a person's prior experience of a word is known in detail, as in a priming study: but the fact that an investigator does not know in detail the mass of a person's experiences that influence the identification of a regular word does not mean that that identification is based on knowledge abstracted across those experiences.

I therefore argue that the basis of word identification cannot be deduced from comparing performance on words differing in frequency or regularity, because the person's specific history that makes those words frequent or regular is not known. Finding that performance is correlated with aspects of general structure does not require that the original learning of that domain proceeded through abstraction of that structure, or that performance in that domain is now controlled by knowledge of abstract structure. Instead, the principles of word identification must be inferred from observations about the acquisition of new words and new experiences of old words, in studies controlling not only the structural properties of items but also the specific processing that people apply to those structures and the contexts in which they are presented in training and test. Under those circumstances, as demonstrated by the experiments described in the last three sections, it is clear that memory preserves specific processing experiences, and that later processing is controlled by the similarity of that event to earlier ones, not only in structure, but also in context and the purpose of the encounter. This can be observed whenever the prior experiences are unique and distinctive, so that their effects can be individually assessed. When instead the person has multiple similar experiences, those experiences will exert simultaneous influence on test performance, so that their individual influences are harder to detect. Performance will be correlated with their average properties, although those properties are not directly computed or directly responsible for performance.

Part 2:

Constructive Evaluation

One of the major barriers to producing a unitary account of memory is the idea that remembering involves retrieval of information from memory to consciousness, whereas in nonreflective tasks such as perception and classification, memory's role is to guide processing, without the guiding representations emerging in consciousness. In this section, I will argue that "retrieval" is a faulty metaphor for remembering: that in the act of remembering, just as in perception, memory representations cued by a current stimulus guide further processing of that stimulus without themselves being transported to consciousness. Further, I will argue that the feeling of "pastness" that separates remembering from other memory-supported activities is not a direct product of interacting with memory, but the result of evaluating that interaction. In the next section, I will extend this argument, suggesting that memory-guided processing followed by interpretation of that processing is a very general phenomenon, and is responsible for all manner of conscious states.

The essential quality of remembering is the feeling of familiarity. A person who tries to remember a past event can tell themselves a story about what might have happened. If that story is experienced without a feeling of familiarity, then the person will not feel that they are remembering, even if the story exactly corresponds to the event. In contrast, the experience of a feeling of familiarity in encountering a stimulus is usually sufficient to make people feel that they have prior experience of that stimulus or one like it, even if they cannot recall any details of the event. The fundamental problem in understanding remembering is thus to understand the source of the feeling of familiarity.

Intuitively, a feeling of familiarity occurs when one encounters a stimulus that one has encountered before. By this intuitive understanding, the earlier occurrence establishes a trace in memory: the second encounter activates that trace, and the activation is experienced consciously as a feeling of familiarity. By this account, the possession of a trace of an earlier experience is a necessary and sufficient cause of the feeling of familiarity.

Contrary to that idea, Jacoby and his colleagues argued that the feeling of familiarity is the product of an unconscious attribution process (cf. Jacoby, Kelly & Dywan, 1989). Jacoby and Dallas (1981) showed subjects a list of words, and then showed old and new words in both a recognition and a tachistoscopic identification test. They observed that subjects were more likely to claim to recognize test items that they could identify more readily. They suggested that the subjects were employing a "fluency heuristic" to perform the recognition task. The idea was that the earlier experience of an item facilitated later processing of that same item presented in the same context (the phenomenon of repetition priming). That additional fluency assisted subjects in identifying the words: They also used it as a basis for deciding that they had seen that item previously. Jacoby and Dallas concluded that the latter inference was performed unconsciously: In the context of a recognition judgment, the extra fluency simply felt like familiarity.

Although that original observation was based on correlational evidence, direct experimental evidence for the fluency heuristic was later provided by several investigators. For example, Whittlesea (1993) showed subjects short lists of words, each word in a list presented for 67 ms. Following each list, subjects were shown a sentence stem, such as "She saved her money and bought a . . .", and asked to read it aloud. The last word of the sentence was then presented, "LAMP" in the example. The subjects named this word, and were then asked whether that word had been shown in the preceding list.

The subjects could discriminate new and old words to some degree, claiming to recognize repeated words 16% more often than new words. They also pronounced repeated words about 100 ms faster than new words. The question was whether that difference in processing fluency produced a feeling of familiarity, a feeling that the subjects could use to perform the recognition task.

Unknown to the subjects, half of the sentence stems were designed to facilitate pronouncing the target word. For example, one stem was "The stormy seas tossed the . . .", completed with the terminal word "BOAT". Compared to the first example, this stem is to some degree predictive of the target word. That covert manipulation reduced the latency of pronouncing target items: subjects were about 130 ms faster in pronouncing target words presented in a predictive context. That is, the test context affected the fluency of naming in much the same way as did a prior presentation of the stimulus. If extra-fluent performance creates a feeling of familiarity, and if people cannot discriminate fluency resulting from a prior exposure from fluency caused by the predictive test context, then manipulation of the test context should produce false feelings of familiarity.

On trials when the target word was presented in the preceding list, the subject could claim to recognize the target either through a feeling of familiarity alone, or by using familiarity coupled with production of some contextual detail of the earlier exposure. However, on trials when the target stimulus was not presented in the preceding list, claims of recognition could only be based on a feeling of familiarity: there was no earlier event to recall. On those trials, the subjects claimed to recognize target words in predictive contexts 18% more often than words in nonpredictive contexts. Apparently, the subjects were sensitive to the fluency of their test performance, and used differences in that fluency to perform the recognition judgment. However, they were not sensitive to the source of the fluency, mistaking enhancement of fluency that was due to the covert manipulation of test context for fluency due to a prior presentation.

I am not suggesting that people usually make errors in feeling objects to be familiar: ordinarily, test contexts are not engineered to provide a spurious source of fluency. However, the illusion of familiarity caused by enhancing the processing of novel test items demonstrates that the feeling of familiarity is not a direct result of cueing a prior experience of the stimulus. Instead, the ordinary development of a feeling of familiarity is a two-step process, a production followed by an evaluation. In the first step, on presentation of the stimulus, cued memory representations construct a percept of the stimulus, and perhaps a full identification of its name, depending on the task. This production does not of itself produce a feeling of familiarity: but it occurs with some fluency. This fluency is likely to be greater if the person has previously experienced that same stimulus in a similar context.

In the second step, the person evaluates the fluency of production. From many prior experiences, the person has an expectation about the fluency with which they could expect to name that word on the basis of their general experience of that word in other contexts. If the fluency of performance in the present context exceeds that expectation, the additional fluency will be attributed to a prior encounter with the word in that same context. This evaluation and attribution are not conducted deliberately or consciously; instead, extra-fluent processing is simply experienced as a feeling of familiarity.

More recently, Whittlesea and Williams (1996) discovered that the evaluative process is driven by the surprise value of the fluency of processing, not by the fluency per se. In one experiment, we asked subjects to memorize a list consisting of natural words (e.g., RAINBOW, TABLE, CANDLE), nonwords that are easy to pronounce (e.g., HENSION, FRAMBLE, BARDEN), and hard nonwords (e.g., JUFICT, STOFWUS, LICTPUB). At test, subjects were shown those items plus an equal number of novel items, of all three types. In this test, they were asked to pronounce each item, decide if it was a word or nonword, and then decide if they had seen it earlier. Novel words were pronounced about 150 ms faster than novel easy nonwords, which in turn were pronounced about 400 ms faster than hard nonwords. Thus if fluency per se is responsible for feelings of familiarity, then we could expect most false alarms on the words, followed by easy nonwords, followed by hard nonwords. That prediction was disconfirmed: false alarms for easy nonwords (37%) were much higher than for words (13%) or hard nonwords (9%). We concluded that the false feeling of familiarity was produced by surprise. Pronouncing the words was easy, but the subjects expected it to be: they knew that they had much prior experience of those words. Pronouncing the hard nonwords was difficult, but again, having determined in the lexical decision task that they were nonwords, the subjects were not surprised at their slow pronunciation. However, the easy nonwords were pronounced fairly fluently, although the subjects knew from their lexical decision that they were nonwords, that they had never seen prior to the experiment. Their pronunciation of those items was thus not just fluent, but surprisingly fluent. Normatively, this extra fluency should have been attributed to the orthographic regularity of the items, not to a prior exposure. However, in the context of a recognition judgment, the subjects experienced this unexpected fluency as a feeling of familiarity.

Looking back, we realized that surprise was also the basis of the illusion of familiarity in the Whittlesea (1993) study, just described. In that case, the subjects were unaware of the context manipulation, and so were surprised when a test item in a predictive context was processed with extra fluency. The idea that surprise is the basis of familiarity helps to explain spontaneous feelings of familiarity, experienced when a person is not being interrogated about their prior experience of stimuli. The feeling of familiarity does not occur when encountering factually familiar (and fluently processed) stimuli in expected contexts, for example meeting one's spouse at home or the clerk in the corner store. One knows who those people are, but does not experience any pressing feeling of "I've seen you before". Instead, that feeling occurs when those stimuli are encountered in an unexpected context, for example meeting the clerk on a bus. In such cases, processing proceeds much more efficiently, completely and elaborately than one would expect for stimuli encountered in that context. Given that there is nothing else obvious to attribute that greater fluency to, it will be attributed to prior experience, and experienced as a feeling of familiarity. Moreover, this helps to explain why the feeling of familiarity stops when one is finally able to produce the context from which one knows the person (the corner store): in that case, the fluent processing is no longer surprising, just as it is not in actually encountering that clerk in the store.

There is of course more to remembering than familiarity, although that is essential. As Mandler (1980) argued, recognition can be performed on either of two bases: by a feeling of familiarity alone, or through additional recall of the contextual detail of the earlier experience. The recall of the detail of an event certainly sounds like retrieval. And indeed, if it could be established that the content of mind on some later occasion was co-extensive with the content of some representation of a prior experience, then it would make sense to speak of that as retrieval of the representation.

However, evidence going back to Bartlett (1932) demonstrates that the detail of an experience that is recalled is not the same as the original experience. Instead, it is subject to various inferences, omissions and additions (cf. the variety of distortions of remembering documented by Franks and Bransford, e.g., 1971, and their colleagues). Even in the immediate case, experiments by Lindsay and Read (1994) and Roediger and McDermott (1995) suggest that the act of recall is as inferential and remote from the representations as is the feeling of familiarity, and for the same reason. Given a list of words, such as BED, NIGHT, PILLOW, DREAM, and BLANKET, and later asked to recall the list, people often falsely claim to recall seeing the word SLEEP in the list. Because that word is thematically related to the others, it is easy for the person to generate SLEEP in thinking about the possible words that might have occurred in the list. Once generated as a candidate for recall, the word must be evaluated: did it come to mind because it was in the list, or because it was thematically related to the other words? The only way that the person can decide this issue is through experiencing a feeling of familiarity for the word or for the contextual detail that comes to mind with the word. But as just demonstrated, that feeling is itself the product of an attribution about the ease with which the word can be processed, not an independent source of evidence about the existence of a representation of that word in that context.

Thus familiarity is fundamental to remembering. In an extended act of remembering, a current stimulus and context cause further ideas to come to mind. Those ideas cue production of further ideas yet, until one has a full-blown story in mind. However, each of those component acts of recall consist in principle of the same thing: a current stimulus complex engages memory, which produces a new mental event. If that new mental event is evaluated as feeling familiar, on the ease with which it comes to mind, it will be used as part of a new stimulus complex to stimulate memory into further production. Never in this chain is there any retrieval. At each stage, memory produces a mental event, and that event is evaluated and attributed to some source.4

The attribution of surprising fluency to a particular source, either in the present or the past, is heuristic and inferential: it is also guided by the context in which the decision takes places. For example, in the Whittlesea (1993) experiment described earlier, subjects were asked whether the target word was pleasant or dull prior to judging whether they had seen it in the list. On trials when the target word had been shown in the preceding list, subjects were 12% more likely to call it pleasant than when it had not been (cf. the mere exposure effect: Zajonc, 1980). Similarly, when the ease of identifying the word was enhanced through a predictive context, subjects were 16% more likely to call it pleasant. Both of these effects are illusions: the words' actual pleasantness was controlled through random assignment to conditions. However, they demonstrate that the same functions that produce feelings of familiarity, namely production followed by evaluation and attribution, also control feelings about the present quality of a stimulus: which the person experiences, a feeling of familiarity or of present quality, depends only on the purpose for which the processing is evaluated.

A large number of judgments about the present quality of stimuli have been shown to be affected by a single exposure of an item earlier in the experiment, including brightness OR darkness (Mandler, Nakamura & van Zandt, 1987), clarity (Whittlesea, Jacoby & Girard, 1989), duration (Witherspoon & Allen, 1985), loudness (Jacoby, Allan, Collins, & Larwill, 1988), truth (Begg & Armour, 1991), pleasantness (Zajonc, 1980), understanding (Carroll & Masson, 1992), and knowing (Jacoby, Woloshyn & Kelley, 1989). Whittlesea, Jacoby and Girard (1989) demonstrated symmetrical illusions, such that judgments about past experience were influenced by the clarity of test presentations as well as prior presentation, and judgments of clarity were affected by prior presentation of test items as well as actual clarity. Which feeling the evaluation function will produce appears to depend only on what possible source of influence the subject is aware of: Surprising fluency of production will be attributed to any likely source, causing a feeling of whatever that source represents.

Although I lack direct evidence as yet, I suspect that the evaluation function is also an important basis of performance in tasks such as classification. In another implicit learning study, Whittlesea and Dorken (1993) showed subjects novel legal and illegal stimuli, but told subjects that half were items they had seen in training and half were new, without mentioning the existence of rules. Subjects were asked to pronounce each item, and then judge it as old or new. Because all items were novel, they could only be judged as new or old on the basis of a feeling of familiarity (rather than recall of context). Items possessing well-formed structure were (falsely) judged old on 63% of trials, whereas illegal items were judged old on only 8% of trials. We concluded that subjects were able to pronounce legal items more fluently than illegal items, because of their greater similarity to training items, and that this caused legal stimuli to feel more familiar. In a paired experiment, subjects were instead correctly informed that some test items conformed to the same rules as training items whereas others were illegal. In this classification test, subjects judged legal items to be legal on 64% of trials, whereas illegal items were judged legal on only 27% of trials. We argued that the basis of this effect was similar to that of the recognition effect: subjects produced the names of legal test items with greater fluency than those of illegal items, attributed this greater fluency to legality, and experienced a conscious feeling of goodness. Under the demand to pronounce and classify the items, the mass of prior experience directly controls the integration of stimulus elements into a percept and the production of a corresponding sound: the fluency of this production can be attributed either to familiarity or to structural goodness, depending on the task.

I therefore argue that remembering and performance tasks do not differ in fundamental process. It is not the case that remembering occurs through retrieval of a specific memory trace whereas performance tasks are supported by application of general knowledge. Instead, both remembering and non-remembering performance are supported by the production of responses toward stimuli, including perceptual, cognitive and motoric responses, driven by representations in memory, and by evaluation and attribution of the fluency of that production. One can use the fluency of producing a name for a stimulus object to decide whether the name is correct ("I think that spice is called coriander . . .yeah, that sounds right"), or the fluency of creating a perceptual organization of a stimulus to decide which category a stimulus belongs to ("That one's certainly a wug: it's so wuggish"), or the fluency of integrating ideas to decide whether some statement is true ("I'd guess that kangaroos are poor swimmers"), or the fluency of processing surface structure to decide that some deep concept is understood ("I had no trouble reading the textbook: I must have understood the material"). The only difference between performance in remembering and non-remembering tasks is the source to which fluent processing is attributed, under the control of the current context and task.

A fundamental idea behind the SCAPE account is that consciousness is not in direct touch either with events of the past or even current objects. Instead, the contents of an experience are constructed, built up through the interaction of representations of prior experiences with the current stimulus compound, but also interpreted within the current context. In the last section, I demonstrated how the evaluative function of memory can produce feelings: feelings of familiarity, goodness, pleasantness and so on. However, its role is greater than simply to produce feelings. It can actually interact with the production function, determining the contents of people's consciousness. This is nicely illustrated by the "repetition blindness" phenomenon.

Repetition blindness was first documented by Kanwisher (1987). She showed subjects sentences, presented one word at a time on a monitor, at a rate of 117 ms per word. Some sentences contained a repeated word, for example "When she spilled the ink there was ink all over". In other sentences, one occurrence of the repeated word was replaced by a synonym, for example "When she spilled the ink there was liquid all over". Kanwisher observed that subjects reported both target words at a high rate when they were different words, but that when a word was presented twice in a sentence, subjects often reported only the first occurrence, even though that rendered the second half of the sentence ungrammatical.

To explain this finding, Kanwisher (1987) appealed to the type/token distinction (e.g., Anderson & Bower, 1973). A "type" is the hypothetical representation of a general concept, a node in a semantic network. The meaning of that type is given by the semantic linkages by which it is connected to other nodes in the network. The strength of these connections is a reflection of the frequency with which those ideas have been associated over the lifespan of the individual. The type thus represents general knowledge, abstracted across many experiences of instances of the concept. In contrast, a token is a representation of a particular event in which the concept participated. The token links contextual information, such as time and place, to the relevant type. When a person encounters a stimulus, the corresponding type representation is automatically activated, allowing the person to know the identity of the stimulus. To remember that they had some particular experience of that stimulus, the person must additionally have created a token on the original occasion, and that token must be activated by the contextual properties at test.

Kanwisher (1987; Park & Kanwisher, 1994) applied this dual-knowledge account to the repetition blindness effect. She reasoned that, in sentences not containing repetition, each target word activated its respective type, and also caused the creation of a token representing the occurrence of that word, so that the subject could later report both target events. The first occurrence of a repeated word also activated its type and created a token, allowing the subject to report the occurrence of that word later. In contrast, she argued, formation of a second token for a type is inhibited if it occurs too soon after the first. In consequence, in rapid list presentation, the second instance of a repeated word does not produce a token, so that the subject has no record of that event having taken place, thus explaining the selective "blindness" that she observed.

The type/token account of repetition blindness suggests a tight linkage between consciousness and memory. By that account, the formation of a memory trace is necessary and sufficient for reporting that one has just perceived a word. Similarly, the formation of two separate traces is necessary to report that the word occurred twice. Having two traces enables the person to remember the word twice, and thereby to know it was repeated in the list. Thus, if a word presented twice in a rapid list is reported only once, one can conclude that the person had formed only one representation of the occurrence of those words. Fundamentally, the type/token explanation treats reporting repetitions in rapid lists as a problem of detection, of registering the event at the time it occurs. Mental events occurring after the list are simply read-outs of traces that were formed earlier.

In contrast, Whittlesea, Dorken and Podrouzek (1995) and Whittlesea and Podrouzek (1995) supplied a "constructive" explanation of the effect. We argued that failure to report both occurrences of a repeated word is not due to failure to encode the second presentation at the time it occurred. Both occurrences are encoded, and representations of both experiences drive that word to come to mind after the list. Instead, failure to report the word twice results from failure to interpret the memory-controlled production of that word as evidence that two occurrences had been encoded. By this account, the relationship between memory representations and mental events is less direct than that suggested by the type/token account. When cued, memory can cause thoughts to come to mind, but these mental events are not necessarily isomorphic with the representations in memory, either in number or content. The occurrence or failure of a conscious mental event is thus not a direct indicator of the past, even of the immediate past. Instead, the person is in the position of trying to make decisions about the nature of their past experience, based on the evidence of what comes to mind now. That is, according to this account, the subject does not have direct access to memory contents, but instead must judge what they just experienced on the basis of their current performance.5

We argued that, during the presentation of a rapid list, each occurrence of a repeated word is encoded in much the same way as a nonrepeated word presented in the same list location. However, items in rapidly presented lists are not processed extensively: Neither is the association between any item and its context (its adjacent words). In consequence, the representations of the two occurrences of a repeated word do not contain much distinctive information. Thus, even if both occurrences are encoded, and both later produce a metal event when the person attempts to recall the list, the mental events they produce will often not be distinctive enough for the person to interpret them as being produced by different sources. In consequence, the person is likely to decide that they saw that word only once, even though their recall of that word is supported by two separate representations. Moreover, the first occurrence of the repeated word, presented earlier in the list, is likely to be more extensively processed with its context. In consequence, people are more likely to be able to produce that context than the context of the second occurrence, and will therefore more likely report that word as having occurred in the first context than the second.

Whittlesea and Podrouzek (1995) demonstrated that reporting repetitions is based on the same attributional mechanism as familiarity. The procedure was almost identical to that of the experiment by Whittlesea (1993) reported in the last section, in which the fluency of pronouncing a target word was sometimes enhanced through a predictive context: The only difference was that target words were presented in the list either once or twice, rather than presented or not presented. Repeated presentation of a word within the list enhanced the latency of pronouncing that word in the later test sentence; so did presentation of the target word in a predictive context. Subjects were 15% more likely to claim that a target word was repeated in the previous list if it was, showing some ability to perform the actual discrimination. However, the subjects were also highly influenced by the enhanced fluency of processing targets in predictive test contexts: On trials when the target word was shown only once, the subjects claimed to have seen it twice 19% more often when it was presented at test in a predictive context than a nonpredictive context. This illusion of experiencing repetition demonstrates that the feeling of repetition is not directly produced by the possession of multiple traces, but is instead mediated by an attribution of the fluency of performance to a source that makes sense. This experiment of course does not directly explain the "repetition blindness" effect: It actually shows the opposite effect, of reporting nonexistent repetition. However, it demonstrates that the logic on which the type/token explanation is based, that reports of occurrences of stimuli are directly related to encoded representations, is false.

Whittlesea, Dorken and Podrouzek (1995) investigated the formation of representations of repeated items. We presented sentences containing a repeated word, such as "When she spilled the ink there was ink all over". We also presented two control conditions, each missing one word required by syntax and meaning: They differed in that the word was omitted either from an early location, called C1 ("The red passed our car on the left") or a later location, called C2 ("The boy hit the little in the playground"). If the contents of consciousness truly reflect the contents of memory immediately after an event, then we should observe that subjects report nonrepeated words (the control conditions) accurately. Further, if memory does not encode a second occurrence of a word occurring soon after the first, we should not be able to predict reports of repeated words from performance in the control conditions, in which repetition-induced failure of encoding would not occur.

First, we found that when subjects reported repeated words only once, they tended to report them in C1 rather than C2, just as observed by Kanwisher (1987): words were reported only in C1 24% more often than only in C2. Such asymmetric report is the basis of Kanwisher's claim that subjects selectively report the first occurrence of a repeated word. However, when we examined the control conditions, we observed the same phenomenon. When a word was omitted early in the sentence, subjects often reported the target that had been presented in C2 erroneously in C1 (producing "The red car passed our on the left"). However, they rarely made the reverse error of reporting a word presented early instead in the late location. Averaged across the two control conditions, subjects were 23% more likely to report a target word in C1 than C2, similar to their early-report rate in the repetition condition.

Clearly, in the control conditions, when there were two locations where they could report a word, but only one word was actually presented that could fit those locations, the subjects had a bias to report that word in C1. That bias is easily understood. In reporting the sentence, they came to that location first: the meaning and syntax of the sentence demand a word at that point, they had available a word to fill it, and so they reported the word at that time. Coming to the second location, needing a word again, but having no other word come to mind and having no evidence that that word had been presented twice, they left that location blank. That is, the subjects were not able to simply read out the sentences from active memory representations: Instead, they constructed a remembrance of the sentence, based partly on what words came to mind, but also driven by their experience of meaning and standard syntax. We argued that that same constructive bias could explain the similar effect in reporting one occurrence of a repeated word. If the subjects believed that that word had been presented only once, they would tend to report it in C1, for the same reasons they tend to report once-presented words there. The asymmetry in reporting repeated words only in C1 versus only in C2 is thus not indicative of which occurrence the subjects are reporting: They could be reporting the second occurrence in the first target location.

We also suspected that at such fast presentation rates, processing of the second occurrence of a repeated word might be independent of processing the same word presented in C1, rather than being either facilitated or inhibited by the earlier experience. To test that hypothesis, we attempted to predict performance on trials containing a target word in both C1 and C2 from the trials containing a target only in C1 or only in C2. We knew that processing of words presented in C1 on nonrepeated trials was necessarily independent of processing words presented in C2, because they were different words on different trials. In the simulation, we combined the data of these conditions as though they were independent contributors to reporting a word presented twice on a single trial. For example, the word would be reported only in C1 either if only the first target was encoded, and was reported only in that location, or if only the second target was encoded, but was reported only in C1, or if both occurrences were encoded, but both were interpreted as a presentation only in the first location. We computed similar probabilities for report only in C2 and for report in both locations. The simulated probabilities matched the actual data for repeated trials very closely: average error was 4% in one study, and less than 1% in a second. We concluded that we could predict performance on repeated trials quite well from knowing performance on trials not containing a repetition: apparently, nothing happens on repeated trials that does not also happen on nonrepeated trials. We further concluded that there was no evidence of inhibition of encoding a second occurrence of a repeated word. Instead, the second occurrence appeared to processed independent of the first.

If there is no inhibition of encoding repeated presentations, why do people report both occurrences of repeated words less often than two nonrepeated words presented in one sentence, as observed by Kanwisher (1987)? If the mental events that occur during the act of recall had unique correspondence to representations, as assumed by the type/token argument, then each representation of the repeated word would sponsor a different mental event, and the person could recall that they had experienced two occurrences. However, as I have tried to argue in this chapter, recall is not retrieval of a trace, but instead the production of a mental event, controlled by representations in memory, followed by evaluation of the source of that event. Both representations of a repeated word would cause that word to come to mind after the list. However, to interpret that coming-to-mind of the word as due to two prior occurrences, the word would have to come to mind as two different experiences. But words presented in high-speed lists are often not well integrated with their context. (It is this poor encoding of the contexts of words that permitted the subjects to report C2 presentations of nonrepeated words in C1.) In consequence, the mental event produced by either of these representations is not much different to that which would be produced by the other: Acting simultaneously, they cause the person to remember the word, but to interpret and experience the coming-to-mind of that word as being produced by a single prior event. In contrast, presentation of two different words in C1 and C2 does not present this problem. After the list, the two representations cause two different words to come to mind, so that the person can conclude they observed two occurrences.

Whittlesea & Wai (in press) provided further evidence that the report of words in rapid sentences is based on construction of a remembrance rather than simply reporting encoded occurrences. In one study, we presented target words twice in a sentence, once in a syntactically and meaningfully congruent context, and once in an incongruent context. The incongruent occurrence always occurred first, e.g., "The diamond ring gift was a gift from his heart". In control sentences, the first occurrence of the repeated word was replaced by a completely different word, e.g., "The diamond ring sock was a gift from his heart". We observed the "repetition blindness" effect: the subjects reported both target non-repeated words on 22% of trials, but both occurrences of repeated words on only 7% of trials. However, target words of both kinds were reported much more often (60% more often) in the second than the first presentation location. Moreover, the deficit in reporting repeated words twice occurred selectively through subjects failing to report the word in the first location rather than the second. This "reverse repetition blindness" effect demonstrates that when people report a repeated word only once, the location of that report is controlled by their understanding of the meaning and syntax of the sentence: They reconstruct the sentence, using available evidence about what words occurred and how often, and where they make most sense. In another study, we provided congruent contexts for all occurrences of target words, whether repeated or not. The manipulation was of the distinctiveness of the contexts of first and second target words, for example, "The teacher gave the test (quiz) today and the test was easy" (less distinctive) versus "He failed the history test (quiz) but the math test was easy" (more distinctive). Using the less distinctive contexts, we observed the standard "repetition blindness" effect: Subjects reported both nonrepeated targets 25% more often than both occurrences of repeated targets, and selectively reported repeated targets in the first presentation location when they reported them only once. When contexts were made more distinctive, reports of the nonrepeated target in the second location increased by 2%. In contrast, reports of the repeated word in that location increased by 15%, reducing the "blindness" effect by half, from 25% to 12%. We concluded that, in reconstructing the sentences, the subjects were more likely to interpret the fluent coming-to-mind of a word as indicating repeated experience when the syntax and meaning of the sentence more clearly supported a second occurrence of the word. These studies demonstrate that repetition detection, like longer-term remembering, occurs through the preservation of the repeated presentations, as they were processed at the time, followed by reconstruction of the experience, based on what comes to mind later, with what fluency, and in what context.

In general, memory drives performance very efficiently, allowing people to deal with the outside world with robust accuracy. However, in making judgments about their own experience, people can only interpret their performance toward stimuli, under some theory about how they would behave in this context if their performance were controlled by this source or that: they have no direct or private access to the source of their behavior. Their beliefs about the nature of their experience are inevitably interpretations and attributions, not direct knowledge. In consequence, those beliefs are highly malleable, subject to modification through changes in the meaning or form of the test context.

General Speculation

The SCAPE theory, based on the selective production, evaluation and preservation of specific experiences, can handle phenomena in a wide variety of areas, through a small set of assumptions. I suspect that it is capable of doing a good deal more work yet. For example, I suspect that it can explain the phenomena of attention, emotion, motivation and consciousness, without any increase in the fundamental assumptions.

The constructs of memory and attention have been dealt with quite separately in the history of psychology, just as remembering and category learning have been. The topic of attention is the selective control of current processing, whereas memory has been treated as the residue of that experience, the storage of whatever was computed when attending a stimulus. Attention is the spotlight, selector or gateway to mind: memory is thought of as the set of knowledge resources that can be drawn on to process selected aspects of the stimulus environment.

I recently witnessed a compelling demonstration, given by Cavanaugh (1995). Cavanaugh pointed out that one of the basic phenomena of attention is to dilate time. To illustrate this, he showed a blinking blue dot, projected on a wall, turned on and off at an SOA of about half a second. After some cycles, the dot unexpectedly expanded until it covered the entire wall. Subjectively, this presentation appeared to be much longer than the other presentations, although it was actually of the same duration. Cavanaugh explained that movement captures attention, and attention dilates subjective time.

Such demonstrations seem to call for an explanation of attention as a function of mind separate from memory. On returning from the conference, I replicated Cavanaugh's demonstration on my laptop. However, in addition, I programmed it in reverse: a dot expanded for a number of cycles, and was then presented as a static display. The reverse phenomenon occurred: the static dot now appeared to endure longer. Moreover, on repeating either of the demonstrations, the subjective dilation of time no longer occurred. This suggested to me that motion is not the key component in the phenomenon of dilated subjective time, and that it is unnecessary to appeal to an extra set of attentional principles to explain it. Instead, I interpret the phenomenon as one of learning and evaluation.

Earlier, I described a study by Whittlesea and Williams (1996), in which subjects were shown words (e.g., RAINBOW), as well as easy (HENSION) and hard (STOFWUS) nonwords. In that study, we observed large false-familiarity effects for easy nonwords, which we interpreted to mean that subjects were surprised by their fluency of processing those items, and, in the context of a remembering task, attributed that unexpected fluency to familiarity. In another study, we switched to examining duration judgments. There was no training phase: but the three kinds of stimuli were presented for either 100 or 200 ms, and post-masked. Subjects were asked to pronounce each item, and then say whether it had been presented for a short or a long duration. Long presentations were judged to be long about 20% more often than short presentations, regardless of stimulus type. However, in addition, short presentations of words and easy nonwords were judged long 15% more often than short presentations of hard nonwords. That is, the subjects suffered an illusion of increased duration with the more fluently processed stimuli.

I do not think that one would want to try to explain this finding in terms of differential attention to the stimuli. Instead, we concluded that the subjective experience of duration is the product of an attribution about the fluency of processing, an inferential process that attributes surprising fluency to a source suggested by the test context. I suggest the same basis for Cavanaugh's (1995) demonstration. The first time a subject watches it, they learn from the induction series of static dots to expect further static dots. Moreover, because the dot is static, it can be completely processed through the production of a single percept. The expanding dot is unexpected. Moreover, because it keeps changing, the subject keeps processing it; and because it expands smoothly, that processing is fluent. This surprising fluency is attributed within the interpretive context to increased duration, rather than to other possible sources, such as familiarity. On a second pass through the demonstration, the subject knows to expect an expanding dot. No longer unexpected, the processing of that stimulus is no longer experienced as taking more time.

I am not trying to downplay the interest of attentional phenomena. Instead, I am attempting to integrate the issues of that area with our growing understanding of memory as an active agent in stimulus processing. I suspect that most, if not all, of the phenomena of selective and divided processing can be understood through the constructs of learning and evaluation, without appealing to a separate set of attentional principles. The phenomenon of negative priming (Tipper, 1985) provides a good example. In that paradigm, a person ignores one stimulus while attending another; for a short while afterwards, the ignored stimulus is harder to identify than is a novel stimulus. This phenomenon initially seemed to require an explanation in terms of inhibitory processing of the ignored stimulus (e.g., Bjork, 1989; Neumann & DeSchepper, 1992), an attentional control process quite unlike learning processes. Like abstraction of general structure, inhibitory processing is a conceivable mental activity: but, to avoid the unnecessary proliferation of entities, before accepting inhibitory processing as real, we should investigate whether negative priming might not be the result of some other kind of process.

The inhibitory-processing explanation rests on the assumption that knowledge about familiar words and objects resides in generic representations in a semantic memory system, representations that have been formed over many prior experiences, and which are little changed by one further experience. The presentation of a word in a priming trial is thus thought of as simply activating the lexical representation of that word, making its identity available. The same logic dictates that if the identity of the word is not available after the word is presented but actively ignored, then the lexical representation must have been inhibited. However, according to the SCAPE account, there is no generic representation of a word to be activated or inhibited: knowledge of a word is distributed across hundreds or thousands of specific prior experiences. Further, those separate experiences are not uniformly cued in a further encounter with a word. Instead, as demonstrated under Theme 4, specifically similar experiences of a word can outweigh the mass of prior experiences. Moreover, every experience of a stimulus is a learning event, in which the specific activities performed on a stimulus in a specific context are recorded. This provides an alternate interpretation of the negative priming phenomenon, that it is a learning effect, as suggested by Williams (1996 ).6 According to this alternate approach, the priming presentation, in which one item is attended and another ignored, is a learning trial, on which the subject learns to produce the identity of one stimulus within a context but also learns not to produce the identity of the other. The probe presentation is a transfer trial. As on any other occasion of encountering a stimulus, memory traces are cued by context, and impose operations on the current stimulus. In this case, one of the prior experiences that is very likely to be cued is the preceding prime trial, on which the person learned not to name that stimulus. That processing is recapitulated in the probe trial, resulting in difficulty in producing the name of the stimulus. Negative priming can thus be understood as a learning phenomenon, a side-effect of constructing and preserving particular experiences, rather than as the product of separate attentional processes.

In general, I think that the study of cognitive psychology is suffering from mis-applied notions of modularity. Clearly, mind can be thought of as serving a number of separate functions, including perception, attention, learning (about both events and concepts), storage, remembering (short- and long-term), reasoning, and awareness. To a great extent, psychologists have elected to study each of these functions in isolation from the others. In doing so, we have also drifted into thinking about them as existing in self-contained units, trading products with each other, but performing qualitatively different activities through qualitatively different mechanisms and by different principles. This drive toward modularity represents good engineering science, as practiced by Western engineers: each part of the machine performs one and only one function, its unique contribution to the overall function of the system. Modularity also makes systems more understandable. If the perceptual system is independent of the remembering module, then they can be studied independently, without worrying that one's conclusions about remembering are only true in some specific perceptual modality.

The drive toward modularity has been fueled by the observation of dissociations between tasks. The idea is that if two tasks are dependent on the same memory representation, then performance in one should be predictable from performance in the other. In consequence, dissociations between tasks can be taken as evidence that they rely on different kinds of representation. For example, Tulving, Schacter and Stark (1982) observed that subjects' performance in an indirect test of memory, such as completion of word fragments, cannot be predicted from their behavior in a direct test, such as recognition, that is performed on the same items. This approach has produced the strongest reason to think of memory as consisting of separate systems, operating by different principles on different aspects of learning and performance (e.g., Schacter, 1987; Tulving, 1983, 1985). Our neuropsychological colleagues have eagerly assisted in this enterprise, supplying physiological correlates to back up the performance dissociations observed by cognitive psychologists (e.g., Squire, 1992; Knowlton & Squire, 1994; Weiskrantz, 1987).

The problem is that there are too many dissociations. For example, Jacoby and Dallas (1981) also observed a dissociation between direct and indirect tests of memory, but their indirect test consisted of tachistoscopic identification rather than fragment completion. Tulving, Schacter and Stark (1982) explained that dissociation in the same way, concluding that recognition is served by one mechanism (episodic memory), and fragment completion and tachistoscopic identification by another (later, two others: the semantic and perceptual representation systems). Then Witherspoon and Moscovitch (1988) showed a dissociation between tachistoscopic identification and fragment completion of the same stimuli. They pointed out that Tulving's dissociation logic requires that flashed versus incomplete presentations of stimuli are processed by different memory subsystems. Hayman and Tulving (1989) even found a dissociation between two successive fragment completion tests on the same items, if the subjects were shown different fragments of each item in the two tests. Tulving and his colleagues currently argue for 5 systems and 12 subsystems within memory, each performing specific, limited functions (e.g., Schacter & Tulving, 1994; Tulving, 1995). There is no obvious end to the number of dissociations and consequent splitting of systems into subsystems, each new pair smaller and more limited in function than the one they replace. Conceptualizing memory as composed of separate functional systems thus seems to be of diminishing utility.

Instead, I am impressed by Jacoby's (e.g. 1983; Jacoby & Witherspoon, 1981) and Roediger's (e.g., Roediger & Challis, 1992) interpretation of the array of observed dissociations. Each task that a person can be set requires some specific resources. Perceptual tasks require prior experience of the perceptual aspects of a stimulus; conceptual tasks require conceptual experience; selective remembering tasks require the encoding of distinctive information about particular events. Any specific experience of a stimulus will involve some combination of these types of information. Any later experience of that same stimulus will benefit to differing degrees from the earlier experience, depending on the type of later task and the resources it requires, in interaction with the resources made available by the specific nature of the earlier experience. This account suggests that the number of observable dissociations is limited only by the imagination of experimenters to contrive combinations of tasks.

I agree that the modular approach is a practical tool for splitting off manageable hunks of the psychological system for study. However, I argue that the various mental functions that psychologists have identified are not real, separate and endemic properties of mind, but rather convenient categories that we have imposed on mind in trying to understand it. As demonstrated by the discussions earlier, I am attempting to take the opposite approach, treating mind as a fundamentally unitary whole.

Perhaps the most important difference between my approach and that driving the development of separate-systems accounts is our assumptions about the directness of the relationship between a cognitive function and the mechanism that serves that function. The idea of separate systems is founded on the assumption that each function is directly served by a mechanism dedicated to that function. Thus, if people are sensitive to abstract prototypes, they must have a prototype-abstraction mechanism; sensitivity to the frequency of words must mean they have a frequency-counter (the strength of a logogen); failure to report second occurrences of rapidly presented stimuli must mean there is an inhibitory mechanism suppressing encoding of repetitions. In contrast, the SCAPE theory suggests that mind performs most of its functions indirectly, almost by accident: All that it ever really does is to construct an experience of a stimulus, under the interactive control of the current stimulus complex and the mass of prior specific experiences, and preserve the new experience. The various sensitivities of the system to abstract properties of experience, although extremely valuable to the person, are side-effects, produced by the distributions of properties across prior experiences, and the cue properties of the current situation. The system can perform the function of identifying a stimulus, producing its name: but it does not do that through an identification mechanism. Instead, the stimulus complex, including the demand to name, cues prior experiences of producing names for objects that look similar to this object, directing the system to impose the same processing on the current stimulus. It can also perform the function of remembering, although it has no retrieval mechanism. Instead, given the task of remembering as part of the stimulus compound, it constructs images of contexts. It can also experience familiarity, by interpreting fluent processing as due to a source in the past; but the same mechanism serves the functions of classification, judgments of temporal duration and feelings of pleasantness, depending only on the task context in which the fluency is experienced.

Memory is fundamentally very simple. The complexity of human performance derives, not from the architecture or processing of memory, but from the variety tasks, stimulus structures and contexts to which memory is exposed.

References

Author Notes

Bruce W. A. Whittlesea, Simon Fraser University. Correspondence concerning this article should be addressed to Bruce W. A. Whittlesea, c/o Dep't. of Psychology, Simon Fraser University, Burnaby, B. C., Canada V5A 1S6. E-mail: bruce_whittlesea@sfu.ca.

Footnotes

1) People can obviously perform abstractions on demand: if asked what writing paper, swans and chalk all have in common, people can produce the answer "white". However, that abstraction is performed in response to a particular demand, not computed automatically and unconsciously through the operation of a chronic abstraction mechanism. It is a specific experience of three stimuli, which will be recorded by memory, like any other experience of the same stimuli for any other purpose. Also like any other event, it will continue to exist in memory as a potential to perform a similar activity on similar stimuli in the future.

2) I am unsure whether to suggest that evaluation inevitably accompanies every production, or only productions that are surprising or questioned by an outside agency. I do not yet fully understand the mechanism through which a person is spontaneously surprised by processing they have just performed. It may turn out that there is a chronic evaluation of every act of production, which only becomes obvious on those occasions on which processing produces unexpected results; alternatively, the mechanism may only act on the latter occasions.

3) Tulving, Schacter and Stark (1982) were forced to go beyond the two stores, of semantic versus episodic information, that had previously been accepted. Jacoby and Dallas' (1981) observation that a single, extra presentation of a well-known word in the training phase of an experiment could influence identification of that word in test could not be explained by either semantic or episodic memory. Tulving's (e.g., 1985) solution was a third memory system, the Perceptual Representation System. Further observations led Tulving (e.g., 1995) to increase the number of proposed memory systems to five. As argued by Witherspoon and Moscovitch (1988) and Roediger (1990 ), there is no reason to suppose that the eventual number of hypothetical memory systems will remain manageably small, because dissociations between tasks is the major indicant of a new division of memory, and because dissociations are being increasingly observed.

4) The production and evaluation of mental events is reminiscent of Anderson and Bower's (1972) generate-recognize theory. However, the underlying assumptions about representation and processing are quite different in the two accounts. Anderson and Bower assumed that knowledge of concepts and events is preserved in an associative network, and that generation and recognition occurred through the spread of activation through the branches of that network. The current article disavows such a global organization of memory.

5) I also argue that the person is in that same position even when a stimulus is physically present. In looking at a word like BUSINESS, I have the idea come to mind that it is the word that means >business<. I can then say that I know the meaning of that word. However, this claim to knowing is inferential, based on the ease with which that meaning comes to mind. This "knowledge" may be spurious: on seeing the word "entomology", the meaning "study of the origins of words" may spring to my mind, driven by the phonological similarity of that word with "etymology", creating the false feeling that I know that word. The fact that the meaning of words usually comes to mind easily and accurately does not mean that one is in direct contact with the memory base: It simply means that masses of prior experience with that word cause its meaning to come to mind fluently when it is presented. Memory drives performance efficiently and generally accurately. However, any reflection on the source of that performance is inferential.

6) Neill & Valdes (1992) also describe a learning account of negative priming. However, their account includes the construct of inhibitory processing. Williams' (1996) account explains the effect simply as an example of transfer-inappropriate processing.