Introduction

Spelling a word, riding a bike, solving a math problem, remembering a shopping list, and finding one's car keys are all memory phenomena. However, even a cursory consideration of these various tasks suggests that they differ fundamentally with respect to key characteristics, such as the level or type of conscious awareness required and the need for contextual support during recovery. Indeed, it is now widely accepted that memory is not a unitary phenomenon, but instead depends on partially independent memory systems, each tuned to fit an environmental demand, each having different operating characteristics, and each reliant on different underlying neural architectures (Schacter & Tulving, 1994; Squire, 1992a; Tulving, 1983).

Empirical research arriving at this conclusion stems primarily from behavioral, neuropsychological, and nonhuman primate investigations; yet arguably the most powerful impetus underlying the multiple systems hypothesis has been reasoned introspection. For example, when recounting his own early thinking on the topic, Tulving (1983, 20) noted:

Another “discovery” had to do with the relation between the learner's response and the internal cognitive state [emphasis added] that it represented. Identical responses could reflect different kinds of awareness. Thus for instance, making a free-association response, such as “chair,” to a stimulus word, such as “table,” does not mean the same thing as the same response made to a question such as, “What was the word that appeared beside the word ‘table’ in the list you just studied?”

The latter question specifically involves the observer's personal history, requiring recovery of information identifying an individual personal episode. Such recovery requires that the observer reflect on his or her past from a first-person perspective. In contrast, the free-association query would make sense to, and presumably would be answered similarly by, large numbers of randomly selected individuals, all with different personal histories. This represents one of the defining phenomenological differences between episodic remembering and semantic knowing, with the latter reflecting the ability to report the properties and regularities of the world that constitute the amalgam of personal experience but no one particular instance. Indeed, it is perhaps safe to say that not a single reader of the current text can recount a particular episode in which he or she learned that in a free-association context, an appropriate response to “table” is “chair”; nonetheless, most readily express this knowledge.

The introspective differences between remembering an episode and reporting semantic knowledge led Tulving (1983) to propose that they rely upon different memory systems associated with distinct types of conscious experience. If that is correct, the relative contributions of the systems could be measured simply by having subjects report their subjective experiences during recognition. This assertion led to the development of the Remember/Know procedure (Tulving 1985), whereby subjects qualify recognition by whether they retrieve a first-person experience of previously interacting with the test probe (Remember response), or whether they instead believe they have recently encountered the probe, but lack specific recollections regarding the encounter (Know response). Use of this procedure has demonstrated principled dissociations between remember and know reports across a range of encoding and retrieval manipulations (Gardiner & Java, 1993a, 1993b) consistent with the idea that they index different underlying memory systems. However, despite the fact that the pattern of Remember/Know results is compatible with a memory systems distinction, considerable controversy exists regarding whether such patterns necessitate multiple systems or can be more parsimoniously explained by a single retrieval system (Dobbins et al., 2004; W. Donaldson, 1996; Wixted & Stretch, 2004).

Although dissociations using subjective reports can be quite compelling, more objective methods also suggest dissociations between episodic and other potential memory systems. One such domain is the study of individuals with damage to medial temporal lobes (MTL) resulting in anterograde amnesia (Cohen et al., 1985; Scoville & Milner, 1957; Squire, 1992b). What has arguably been most remarkable about amnesic patients is not so much their global and prominent impairment in learning facts and remembering events, but the preservation of an immense range of memory skills (Moscovitch et al., 1994; Schacter & Graf, 1986; Tulving et al., 1988; Warrington & Weiskrantz, 1974). For example, the patient E.P., who suffers from extensive bilateral MTL damage, demonstrates a virtually complete amnesia for episodic content even when indexed by simple recognition (Hamann & Squire, 1997). Despite this, he is able to provide exquisitely detailed descriptions of the spatial layout of his boyhood hometown (Teng & Squire, 1999) indicating that his long-term semantic knowledge regarding its layout is fully intact. Despite this, he is unlikely to remember recounting this information to investigators even mere minutes later.

In addition, these patients typically demonstrate normal or near normal gains in the ability to successfully complete fragmented word probes (e.g., a_rd__rk) following study of previously intact items (e.g., aardvark), and likewise appear normal on a host of other implicit tests of memory, in which no overt reference to prior study episodes is made (Roediger, 1990; Warrington & Weiskrantz, 1974). Thus, although amnesic subjects are often completely unaware of the study/test relationship, performance nonetheless improves as a function of prior exposure. Findings such as these have led to the postulation of an implicit memory system complementing the episodic and semantic systems alluded to above.

Neuropsychological dissociations are powerful demonstrations of the relative sparing and destruction of different memory systems. However, these hypothetical systems are also dissociable experimentally in healthy subjects. For example, explicit and implicit tests have been shown to demonstrate different forgetting functions (Tulving et al., 1982).

Above we have discussed a mere fraction of the evidence for a systems approach to understanding memory behavior, and have bypassed altogether animal research also supporting similar taxonomies (Zola-Morgan & Squire, 1992). Indeed, even a nominal account of any major area (phenomenological, neuropsychological, cognitive behavioral, animal/comparative, developmental) would itself entail a very large chapter or book, and we point interested readers to several texts and reviews (Bauer, 2004; Foster & Jelicic, 1999; Mayes & Downes, 1997; Squire & Butters, 1992). However, even the cursory review offered above suggests a considerable amount of convergent evidence supporting the idea that human observers are endowed with separable memory systems that differ fundamentally in terms of their operating characteristics, conscious correlates, and behavioral affordances. In this chapter, we will focus exclusively on episodic memory.

Before turning to functional imaging studies of episodic memory, it is important to consider more carefully the behavioral techniques used to measure it and the core assumptions governing those techniques. As emphasized by Tulving's encoding specificity principle (Tulving, 1983), understanding episodic memory requires careful consideration of the interaction between encoding and retrieval conditions (cf. Morris et al., 1977; Roediger, 2000). More specifically, although the tendency for an event to be forgotten critically depends on the manner or circumstances in which it is encoded, consideration of the encoding conditions alone cannot determine the status of an episodic memory. For example, I may ask you whether you met anyone new at a particular event and you may emphatically state that you did not. Such a query is termed a recall or cued-recall task. Following this denial, I may then present a person to you and ask if you met him or her at that same event. During this source recognition task, you may indeed remember the previous meeting and perhaps recover further details about the encounter. This represents a case in which the apparent status of your memory depends heavily upon the format of the retrieval test; the reverse also holds, different conditions during the experiencing of the event critically affect whether memory will be evident on a particular type of subsequent test. Thus, although we later consider encoding and retrieval operations separately, it should be understood that the two are intimately related and, indeed, one cannot be understood without the other.

As noted above in the quote from Tulving, equivalent memory output responses can arise from fundamentally different internal cognitive states. This belies the fact that measures such as accuracy and response time, although vital, have inherent limitations with respect to disentangling underlying cognitive causes. In contrast to early suggestions regarding the substrates of different memory systems, researchers now have access to technology that enables the cataloging of similarities and differences across internal cognitive states (or at least their neural correlates). As we shall see below, although the functional imaging of episodic memory has laid the groundwork for characterizing episodic encoding and retrieval in terms of distributed neural systems, questions remain about the functional roles of cortical regions composing these systems, their interaction, and the compatibility of cortical activity findings with respect to the multiple memory systems framework.

Before turning to a review of the findings of functional imaging studies with respect to episodic memory, it is important to cover one more critical distinction when describing episodic memory (other than encoding versus retrieval), and that is the distinction between processes directly reflecting memory evidence versus those indicative of working with or deciding about that evidence (decision criteria) (Moscovitch & Winocur, 2002). Behaviorally, this distinction is typically formalized using statistical decision models such as signal detection theory (SDT) (Macmillan & Creelman, 1991) (figure 8.1).

Figure 8.1.  

Illustration of the Signal Detection model of recognition. Top panel: The evidence underlying recognition is assumed to be continuous. Old and new items in a recognition test are assumed to be normally distributed around central values separated by a distance (d′). Because the distributions overlap, observers place a decision criterion along the evidence axis in order to parse items into “old” and “new” classifications. This yields four possible responses types (hits, false alarms, misses, and correct rejections). Bottom panel: Two observers (or the same observer on different occasions) can have identical evidence yet yield very different response patterns as a result of a shift in the decision criterion.

During episodic recognition experiments, the primary paradigm used in functional imaging, subjects are shown a list of stimuli, such as words or pictures, during an encoding phase. These items are then re-presented during the retrieval phase, intermixed with new, unstudied items. The observer's task is to discriminate between old and new items. The SDT model assumes that memory evidence is continuous and normally distributed around two central tendencies, one for old and one for new items. Because these evidence distributions overlap, observers must select a cutoff, or criterion that reflects how much evidence is deemed sufficient to classify an item as studied. Whereas the absolute ability of observers to distinguish old and new items is determined by the distance between the evidence distributions (termed d′), the decision criterion determines how that evidence is overtly expressed. Hence, within SDT, overt responding is a function of both raw memory evidence (characterized as d′) and decision factors regulating the placement of criterion (termed c or β). Critically, two observers with identical raw evidence may nonetheless adopt very different decision criteria. Thus, whereas Tulving (1983) emphasized that two people can express identical overt memory responses via fundamentally different internal cognitive states (see above), SDT presumes it is also possible for two people to make fundamentally different memory classifications (“old” or “new”) based on exactly the same memory evidence. A major challenge facing the interpretation of functional imaging data is disentangling decision mechanisms that regulate valuation or monitoring of memory evidence from those which express the raw quantity or quality of that evidence.