It has been clear for decades that the hippocampal system plays a prominent role in learning and memory. Only more recently has it become clear that, within this role, only particular aspects of memory processing depend on hippocampal system function and that these properties can be distinguished operationally from other aspects of memory that do not depend on the hippocampal system. However, reaching an understanding of memory rich enough to characterize the nature of hippocampal-dependent and hippocampal-independent memory processing, and tie them to underlying neural mechanisms, has been difficult and somewhat contentious. The earliest reports of amnesia emphasized as the central feature of hippocampal function its time-limited role in memory, focusing on the critical role of the hippocampal area in "consolidation" processes that bridge between immediate memory and the long term store (Scoville & Milner, 1957). This temporal distinction is most striking: Amnesic patients and animals with hippocampal system damage show normal retention at short delays and increasingly impaired retention with increasingly long delays. More recently, some (e.g. Rawlins, 1985) have argued that virtually all aspects of memory processing by the hippocampal system can be attributed to its function as a temporary memory store or buffer. However, a large body of data indicates that the loss of "consolidation" or memory-buffer functions cannot account for the full pattern of impaired versus spared memory performances observed after hippocampal system damage; some memory performances are intact in amnesia even across very long temporal intervals, whereas other memory performances are grossly impaired even over relatively short retention intervals. This view also fails to account fully for a number of relevant observations on the stable, long-lasting functional correlates of hippocampal neuronal activity. Such findings are considered latter in this paper.
To accommodate these findings many investigators have proposed a second kind of functional distinction of hippocampal system function. Specifically, it has been argued that the hippocampal system plays a role in only one "type" of memory or, more precisely, only one form of memory representation; understanding the nature of this representational distinction of hippocampal-dependent processing has become a dominant theme in contemporary memory research. Significant progress has been made in characterizing the type of memory representation impaired in human amnesia and in the form of memory representation that is deficient in animals with experimental damage to the hippocampal system. Although there continues to be some disagreement about the scope of the deficit and the terminology used to describe it, most investigators agree that there are multiple memory systems with different representational characteristics and that the hippocampal system is critical to only one such system. Based upon our view that the hippocampal system is involved in declarative memory but not procedural memory, in the present commentary we will adopt the terminology that distinguishes a hippocampal-dependent capacity for relational representation, which supports memory for relationships among perceptually distinct items and the flexible expression of memories in novel contexts, from a hippocampal-independent capacity for individual representations, which involves the acquisition of biases and adaptations to individual items expressible only through repetition of the learning event. This will be explicated more fully below (see pp. 18-19).
To date, little research has been directed toward exploring the potential relationships between the temporal and representational distinctions of hippocampal-dependent memory processing. It is probably fair to say that most investigators, ourselves included, have assumed that the time-dependent properties of hippocampal system involvement are merely a qualification of the role of this system in a particular type of memory representation. However, findings derived from a number of behavioral paradigms are more clearly related to the temporal properties than they are to the distinction between forms of memory representation, in a manner that suggests that the temporal properties of hippocampal memory processing are not subsidiary to the representational properties. The present commentary considers hippocampal memory processing in terms of these two orthogonal properties, one that involves the strength and persistence characteristics of storage within the hippocampal system and another that involves the form of memory representation which this system supports. It will be argued that (1) hippocampal-dependent memory differs from hippocampal-independent memory with regard to both these properties, (2) specific findings with the full range of impaired versus spared memory performances in amnesia associated with hippocampal system damage will be understood with respect to one or the other of these two distinguishing properties of hippocampal functioning, and (3) these two properties are mediated by different anatomical components within the hippocampal system. Furthermore, we will argue that the combined temporal and representational properties of the hippocampal system processing compose the fundamental characteristics of declarative memory in humans and animals.
The Plan of this Paper
In the following sections we will describe an account of the flow of information in the hippocampal system, suggesting specific anatomical assignments for aspects of hippocampal memory processing responsible for the temporal and representational distinctions described above. The plan of this paper is as follows: First, we will consider the scope of this presentation and some anatomical designations critical to our model. Second, we will sketch the model briefly to introduce the conceptual issues and make empirical predictions. In the third and fourth sections, we will further characterize and discuss the empirical basis of the temporal and representational distinctions of hippocampal function, respectively, as they are derived from the work on human amnesia and work on animal models of amnesia (emphasizing particularly the animal work); these sections will also outline the evidence that these functions can be dissociated by selective lesions within the hippocampal system. Fifth, we will provide converging evidence on the coding properties of neurons in the cortex and hippocampal system of animals performing memory tasks. Sixth, we will combine the neuropsychological and neurophysiological data in a more detailed description of the model, more fully explicating our view about how these properties contribute to, and interact in, declarative memory. Seventh, and finally, we will compare our model with other theoretical accounts of hippocampal function and, in doing so, will demonstrate how the properties of the model draw together central features of these proposals, thereby providing a more comprehensive account.
1. Some Preliminaries
Before considering the central issues of this presentation, it behooves us to address some general issues regarding the scope of the present account and our set of anatomical conventions.
Limiting the scope of the present account. The aim of this article is to present and support an hypothesis about the distinct and interactive functions in memory of components of the hippocampal system. In order to keep the focus on this topic, some issues will not receive a comprehensive review. We will not consider the large literature reporting effects of hippocampal system damage on behaviors only indirectly related to learning and memory. These include studies on orientation, distraction, exploration, motor patterns, operant schedules, emotion, and species-specific behaviors (for detailed reviews see Gray and McNaughton, 1983; Gray, 1982; O'Keefe and Nadel, 1978). By adopting this limitation, we are assuming that alterations in these behaviors following hippocampal system damage are either a consequence of amnesia or an indirect result of disconnections of the limbic system that have non-mnemonic as well as mnemonic consequences (for further discussion of this issue see Gray and Rawlins, 1986). Regardless of the ultimate disposition of these findings, a discussion of non-memory effects of hippocampal system damage would constitute an major divergence from our main interest here.
Also, our review of the literature regarding the effects of hippocampal system damage on learning and memory will not be exhaustive. We will focus on the literature that is particularly relevant to contrasting temporal and representational distinctions, and to dissociating these properties among components of the hippocampal system. We will not provide a comprehensive review of behavioral evidence from studies on human declarative memory, on human amnesia, or on animal learning and behavioral assessments that are not directly relevant to the two distinctions. Thus, our review will necessarily emphasize the animal literature, where selective damage and recording studies can and have been conducted; overviews of the human literature will be employed only to show how this work has generated the initial observations about the temporal and representational distinctions. Furthermore, even our coverage of the animal literature will necessarily be selective, considering carefully a set of behavioral paradigms that are particularly illuminating with regard to the distinct temporal and representational properties of hippocampal function. A more extensive survey of the literature, particularly with regard to a full analysis of relational representation in human declarative memory and amnesia, can be found in Cohen and Eichenbaum (1993).
Delineating the anatomical components of the hippocampal system. Since there is no universal agreement on just what constitutes the term "hippocampal system", we must define our use of this term here and justify the selective inclusion of specific brain structures within what we will argue are the functional components of the system. The term "hippocampal region" was first used to describe the set of medial temporal lobe structures removed in the patient H.M., whose resulting amnesia has been the subject of neuropsychological study for nearly 40 years (see below). In his case the surgical removal included most of the hippocampus (Ammon's horn), the dentate gyrus, the subicular complex, the amygdala, and, potentially, parts of several cortical structures including the entorhinal, perirhinal, and piriform cortices (Scoville and Milner, 1957). Comparisons of H.M.'s memory performance with that of patients having more restricted medial temporal lobe removals indicated that the degree of damage to the hippocampus per se determined the severity of amnesia; consequently much memory research has focused on this particular structure. However, subsequent efforts to model the amnesic syndrome in animals by selectively ablating the hippocampus were disappointing; monkeys and rats with hippocampal lesions were not severely impaired at many common discrimination learning tasks employed in memory research on animals.
A breakthrough in resolving this discrepancy came when Mishkin (1978) reported that monkeys with conjoint but not separate damage to the hippocampus and amygdala were severely impaired in recognition memory for objects, suggesting that a medial temporal lobe removal as complete as that of H.M. was necessary and sufficient to produce severe global amnesia in animals. However, more recent experiments have modified this conclusion in two ways. First, it is now clear that restricted amygdala lesions do not cause impairment on tasks sensitive to large medial temporal damage, and they do not exacerbate the memory deficit resulting from damage to other medial temporal structures (Zola-Morgan et al., 1989a). These findings redirected attention to the damage caused in cortical areas adjacent to the hippocampus that inevitably occurs during surgical removal of the amygdala or hippocampus. Second, it has recently been demonstrated that substantial damage to the surrounding cortical areas alone produces as great a deficit in certain task performances as does removal of the entire medial temporal lobe, indicating that these areas play a direct and critical role in memory processing (Zola-Morgan et al., 1989c; Murray and Mishkin, 1986; for review see Squire and Zola- Morgan, 1991).
Based on the current state of the neuropsychological findings, we conclude that the critical structures comprising the medial temporal memory system include the hippocampus proper (Ammon's horn), the dentate gyrus, the subicular complex, and surrounding cortical areas including the entorhinal, perirhinal and parahippocampal areas. Furthermore, for purposes that will become apparent, we have adopted a terminology consistent with that suggested by Witter and colleagues (1989) to distinguish the hippocampal formation and the parahippocampal region as the two major functional components of the hippocampal system. In the following discussion we will include within the hippocampal formation Ammon's horn, the dentate gyrus, the subiculum, and the fornix; and within the parahippocampal region the entorhinal, perirhinal, and parahippocampal cortices. [Note: There is a danger of confusing the terms "parahippocampal cortex", which is discrete area in the primate cortex, and "parahippocampal region", which encompasses this specific cortical area plus entorhinal and perirhinal cortex.] A schematic diagram indicating our functional assignments for these areas and the flow of information between these areas and the neocortex is provided Figure 2. In this summary we introduce the fundamental anatomical associations between the neocortical areas that provide specific perceptual and motor information to the hippocampal system, the parahippocampal region which serves as a convergence center for neocortical inputs and mediates two-way communication between cortical association areas and hippocampal processing, and the hippocampal formation itself (Deacon et al., 1983; Van Hoesen et al., 1972; Amaral and Witter, 1989).
Before addressing in detail the functional assignments proposed for these areas, it is important to justify this particular anatomical segregation of functional units within the hippocampal system. With regard to the hippocampal formation, each of the subdivisions of this area are well defined and largely comparable in the monkey and rat (Figure 1; cf. Rosene and Van Hoesen, 1988). Our conclusion that the subiculum should be included among the structures that comprise the hippocampal formation is supported by both neuropsychological and anatomical findings. The most relevant neuropsychological evidence comes primarily from the elegant experiments of Jarrard and colleagues that employed selective neurotoxic lesions to damage particular subdivisions of the hippocampal formation (see Jarrard, 1986, 1993). Results from these studies consistently indicate that ablation of all of these components is critical to producing severe memory impairment in spatial and non- spatial tasks (Jarrard, 1986). Notably, Jarrard has interpreted his findings as indicating a preferential role for Ammon's horn and the dentate gyrus in spatial learning. This conclusion was based largely on the finding that damage to Ammon's horn and the dentate gyrus are sufficient to produce spatial memory deficits and that subicular damage must be included to produce a deficit on nonspatial memory. However, more recent work using the sensitive test of the Morris water maze indicates that a severe deficit even in spatial learning is found only after combined damage to the subiculum as well as Ammon's horn and the dentate gyrus (Morris et al., 1990). Thus, based on the current state of the findings, it appears that the most restricted hippocampal system lesion that produces maximal impairment in any task must include the combination of structures that comprise the hippocampal formation. With regard to anatomical evidence, consistent with the behavioral data, Amaral and Witter (1989) and Witter (1989) viewed the subiculum as the final stage of intrinsic processing in the connectional "loops" between the parahippocampal and hippocampal areas. Their reviews of hippocampal pathways placed the subiculum as the fourth and final stage of serial hippocampal processing through the so-called trisynaptic circuit, and drew close parallels between the parahippocampal connections with Ammon's horn and with the subiculum.
In addition, it is important to consider at the outset a designation for the fornix, the major fiber bundle that connects all subdivisions of the hippocampal formation with the septum and other subcortical areas. A complete transection of the fornix disrupts cholinergic and GABAergic function and electrical activity, and induces morphological reorganization in the hippocampal formation (e.g. Lahtinen et al., 1993). Thus a fornix transection produces significant disruption of information processing and output of the hippocampal formation, and so would be expected to have the same functional consequences as damage to the hippocampal formation itself. Notably, a fornix transection does not disconnect the parahippocampal region from the neocortex. Thus a fornix transection may not disrupt functions that can be carried out by the parahippocampal region independent of processing by the hippocampal formation, and hence would not be expected to produce the full-blown amnesia seen following more complete hippocampal system damage. Combining these behavioral and anatomical observations, we will hereafter consider the subiculum and fornix, along with Ammon's horn and the dentate gyrus, as critical elements of a functional unit within the hippocampal system.
With regard to the parahippocampal region, all three cortical areas that we include within this region are well defined in the monkey, but less so in the rat (Amaral et al., 1987; Witter et al., 1989, Deacon et al., 1983). In both species the entorhinal cortex has been characterized as a cytoarchitecturally distinct area surrounding the hippocampal formation and contributing the major cortical afferents to the hippocampal formation. It is a heterogeneous area composed of several subdivisions with topographically contiguous and often overlapping projections into the hippocampal formation (see Amaral and Witter, 1989). The perirhinal cortex, composed of Brodmann's areas 35 and 36 bordering the rhinal sulcus, is also distinguishable in both the rat and monkey, although the extent of its borders in the rat are not fully clear. The parahippocampal cortex in the monkey is clearly distinguished as the combination of von Bonin and Bailey's areas TH and TF lying posterior to the perirhinal cortex. It is not generally agreed whether the same area exists in rats, but connectional correspondences have led Deacon et al. (1983) to propose that the 'postrhinal' cortex that forms the caudal continuation of rhinal sulcus is homologous to the primate parahippocampal cortex. In addition to these cytoarchitectural distinctions, the subdivisions of the parahippocampal region are distinguished from each other by several connectional differences. For example, in the monkey, the perirhinal area receives the greater inputs from higher visual areas while the parahippocampal cortex receives the greater input from the parietal and cingulate cortices (Suzuki & Amaral, 1993, 1994); a similar topography of inputs distinguishes the perirhinal and postrhinal areas of the rat (Deacon et al., 1983). A major similarity of the perirhinal and parahippocampal (postrhinal) areas is that both heavily project to parts of the entorhinal cortex. Despite the various differences in the cytoarchitecture and extrinsic connections of the entorhinal, perirhinal, and parahippocampal cortices, each of these areas can be characterized as receiving input from multiple neocortical association areas and thus comprises an important convergence site for neocortical input to the hippocampus (Squire et al., 1989; Witter et al., 1989; Deacon et al., 1983). Furthermore, each of these areas is reciprocally connected with both neocortical areas and with the hippocampal formation, and thus serves as the major intermediary in communication between the hippocampal formation and neocortex. While the precise pattern of cortical and hippocampal connections differs among these structures, the neuropsychological data have not as yet revealed any corresponding functional distinctions (see below and Witter et al., 1989). Given these commonalities, we will refer to these areas collectively at the risk of overshadowing potentially important functional distinctions that may yet be discovered.
In addition, there is some controversy about the limits of the cortical areas that should be included within the parahippocampal region, particularly as related to visual memory in primates. In primates, just adjacent to the parahippocampal region is a portion of the inferotemporal cortex (anterior inferotemporal cortex or TE1), an area critically involved in higher order visual identification and memory (Gross, 1973). As will be discussed below, some of the memory processing properties of cells in this area are very similar to the features of cells in the parahippocampal region, suggesting at least part of this area may be more reasonably considered as an adjunct to the parahippocampal region, with its role limited to visual memory processing.
Finally, we should acknowledge that there is a danger of circularity in defining the hippocampal system based on neuropsychological findings because, as will become clear in the following discussion, a precise characterization of the memory processes supported by this system is itself still evolving. Thus, the operational definition of "the hippocampal system" may change as we refine the characterization of memory deficits resulting from circumscribed damage to particular structures. For the moment, however, we interpret the converging anatomical, behavioral, and electrophysiological evidence as suggesting particular temporal and representational functionalities mediated differentially within separate components of the hippocampal system.
2. The Role of the Hippocampal System in Intermediate-term Storage and Relational Representation: A Model of Successive Stages in Hippocampal Mediation of Declarative Memory.
Below we propose a model that attributes to the hippocampal system two sequential functions corresponding to anatomically separate components of this system. First, the hippocampal system has the capacity to represent isolated (non-relational) items at full strength and to hold these representations in a memory "buffer" for periods of at least several minutes; the persistence of these representations may depend on the nature of the stimulus material and can be extended considerably by repetition (see discussion on p. 16). Intermediate-term memory is thus defined here as bridging the gap between the very brief span of immediate (or short-term) memory and the potentially permanent (or long-term) memory store. Unlike immediate memory and like permanent memory, the intermediate- term store is viewed as sustaining retention across some interfering or distracting event but, unlike permanent memory, this form of representation cannot be sustained indefinitely. Second, during this intermediate period, the hippocampal system subserves processing that involves comparing and relating these individual representations to other memory representations, creating relational representations according to the relevant contingencies between the items and the structure of any already established memory organization that involves those items. The combination of these two processing functions constitute the properties of declarative memory as we have characterized it.
According to this account, while these two processing functions are supported independently, they normally function interactively, with the relational memory processes operating upon an intermediate-term store of new items. Thus the temporal and representational properties of the hippocampal system are considered partially interdependent; relational memory processes depend on the intermediate-term store, but intermediate-term storage of single items does not require relational memory processing. This hypothesis is most simply instantiated by a two-stage model in which the intermediate storage function is accomplished at an earlier stage of the hippocampal system than is relational processing. In the present paper we propose that just such a functional distinction exists, specifically that intermediate-term storage of individual representations is supported primarily by the parahippocampal region and that the hippocampal formation mediates relational processing (Figure 1). These two subdivisions of the hippocampal system will be distinguished from the closely related neocortical association areas that are reciprocally connected with the hippocampal system and, as will be argued below, contribute to memory processing in different ways.
Our proposal regarding the functional roles of the hippocampal formation and the parahippocampal region generate two sets of testable predictions. First, the present account predicts that damage to the parahippocampal region should result in a loss of intermediate-term storage capacity for individual items; this damage should also inevitably result in a deficit in relational processing consistent with the view that intermediate-term storage is a prerequisite for a relational memory representation. By contrast, damage to the hippocampal formation or its non-cortical connections via the fornix should selectively affect relational memory processing, leaving intact a capacity for intermediate-term storage of non- relational information. Second, additional evidence in support of the model might be obtained through observations on the behavioral physiology of the hippocampal system and the neocortical structures that are the source of much of the information this system receives. The model predicts that neural activity within the parahippocampal region should reflect intermediate-term storage for specific and isolated items and events, whereas activity in the hippocampal formation should be strongly associated with relations among those items. To our knowledge no existing experiments were explicitly intended to test these predictions. Nevertheless, there exists a large body of data on the behavioral effects of damage to the hippocampal system, including several studies involving selective damage restricted to either the parahippocampal region or the hippocampal formation; many of these studies assess the effects of these lesions on performance that depends upon intermediate-term storage or relational representation. In addition, there have also been a number of recent examinations of the behavioral physiology of both the cortical association areas and the hippocampal system in animals performing recognition memory and other learning tasks. In the following sections we will summarize some of the relevant findings, showing how they support predictions of the theoretical claims that were outlined above and that will be explicated more fully below.
3. Temporal Properties of Hippocampal-Dependent and Hippocampal-Independent Memory.
In 1953 the patient H.M., who for years had suffered frequent major epileptic seizures intractable to pharmacological treatment, was operated upon for removal of the medial temporal lobe area, including the hippocampus and neighboring structures (Scoville and Milner, 1957). Subsequent to this surgery he demonstrated a profound deficit in new learning that apparently extended to all types of materials, leading to the characterization of his impairment as "global" anterograde amnesia. However, from the very first reports it has been clear that the range of this amnesia was distinguished by its restricted temporal domain. H.M. was unable to retain nearly any new information for extended periods, although he demonstrated an intact immediate or short-term memory capacity. In addition, much of his remote memory was also intact, exemplified by his retention of language skills, general world knowledge, and childhood memories. Despite the preservation of remotely acquired information, memories H.M. acquired more proximal to his operation were lost; his retrograde amnesia extends 3-11 years before his surgery (Corkin, 1984). Contrasting H.M.'s deficit in new learning with his intact short- term and remote memory capacities suggested that the amnesia associated with hippocampal system damage could be viewed as an impairment in the encoding, maintenance, or "consolidation" of enduring and accessible memories. The same pattern of temporally- dependent impairment and sparing of memory has since been described in several other patients with medial temporal lobe damage (see Squire, 1987).
Animal models of hippocampal involvement in intermediate- term memory. A considerable number of studies, using a variety of learning paradigms, have demonstrated both delay-dependent impairment in new learning (Alvarez-Royo et al., 1992; Gaffan, 1974; Mishkin, 1978; Zola-Morgan & Squire, 1985; Aggleton et al., 1989; Staubli et al., 1984, 1986; Kesner & Novak, 1982; Otto & Eichenbaum, 1992a; Overman et al., 1990; Staubli et al., 1986) and temporally-graded retrograde amnesia (Zola-Morgan & Squire, 1990; Kim & Fanslow, 1992; Winocur, 1985, 1990) in animals with hippocampal system damage. These findings make it clear that the hippocampal system is neither a critical site for short-term memory, nor the sole final repository for long term memory, but rather plays a critical role during the intermediate period that bridges between them. The locus of storage for short-term and permanent memories is not known, but is widely believed to be in the neocortical areas interconnected with the hippocampal system. Also, whether the hippocampal system actually stores memories or mediates some interaction among other (presumably neocortical) storage sites during the time after learning when hippocampal system involvement is necessary remains unclear. For further discussion on these points, see below and reviews by Eichenbaum et al., (1992a), Teyler & Discenna (1986), Squire, Cohen, and Nadel (1984), and Halgren (1984).
Recent efforts to delineate the anatomical structures involved in intermediate-term memory maintenance have focused on the delayed non-match to sample task (DNMS). Each trial of the DNMS task is composed of three phases. In the first phase a sample memory cue is presented. This is followed by a delay phase during which the memory for that cue must be maintained. Finally, in the choice phase, the subject is presented with the sample and/or a novel stimulus, and the unfamiliar "non-match" cue must be selected. The load on memory can be increased by lengthening the delay phase or by presenting a "list" of sample cues prior to a series of choice recognition tests. This paradigm was first developed for monkeys, using three dimensional junk objects that provide rich and salient cues for this species (Gaffan, 1974; Mishkin & Delacour, 1975). Stimuli were used only once or very infrequently across trials, thereby eliminating the high level of inter-item interference prominent in other related tasks, including delayed match-to-sample with trial- repeated cues, delayed response, and delayed alternation. Damage to the hippocampal system results in a performance deficit that is dependent on the duration of the delay or the length of the sample "list" (Gaffan, 1974; Mishkin, 1978), revealing a delay-dependent deficit in new learning similar to that observed in human amnesics (e.g. Mishkin, 1982). Indeed amnesic patients also demonstrate delay- dependent impairment on versions of DNMS developed for testing human subjects (Squire et al., 1988; Aggleton et al., 1988). Finally, the DNMS paradigm has been revised successfully for testing recognition memory in rats by using visual-tactile or olfactory stimuli as memory cues (Aggleton et al., 1986; Rothblat & Hayes, 1987; Mumby et al., 1990; Otto & Eichenbaum, 1992a) and, in some of these studies (see below), damage to the hippocampal system results in the same deficit as observed in monkeys and humans.
While the DNMS paradigm has been extraordinarily fruitful in studies delineating the components of the hippocampal system necessary for this type of simple recognition memory (see Squire and Zola-Morgan, 1991), it is not clear just how much the cognitive demands of DNMS depend on the representational properties of hippocampal-dependent memory introduced above. The hippocampal system might participate in this kind of recognition memory by supporting a judgement of the relationship between perceptual information available during the choice phase and memories for experiences with previous stimuli. The hippocampal system may also play an important role in the requirement for a non-match response because non-matching requires a "flexible" response, that is, a choice contrary to that performed during the sample phase. For both of these reasons, hippocampal processing may confer an advantage to normal subjects. However, given that DNMS is typically run with trial unique stimuli, hippocampal-independent memory processing based upon the familiarity of sensory features of individual stimuli might support DNMS performance without representation of relations between the sample and other previous stimuli or of the episode surrounding previous experience with a familiar item. Indeed, recent work with human amnesia suggests that "familiarity" or perceptual fluency can provide a basis for some recognition memory judgements (e.g. Verfaillie & Treadwell, 1993). It is notable that the non-match response requirement of DNMS is not critical to the demonstration of a memory deficit; monkeys with hippocampal system damage are significantly, if not as severely, impaired when the requirement is to "match" to the sample cue (Gaffan et al., 1984; Mishkin et al., 1984). Most importantly, between matching or non-matching and across the range of variants on DNMS tasks, the presence and magnitude of impairment after hippocampal damage is primarily determined by the duration of memory delay, or the delay introduced when lists of sample stimuli are used (see, especially, Alvarez-Royo et al., 1992; Overman et al., 1990). These findings suggest that the DNMS task is particularly sensitive to the strength and persistence characteristics of hippocampal memory processing.
The findings on object discrimination learning tasks are relevant and illuminating here because, as with the DNMS task, the performance of animals with hippocampal system damage is best predicted by the duration of memory delay and not by appeal to a representational distinction (cf. Mishkin et al., 1984). Monkeys with hippocampal system damage are mildly impaired in learning simple object discriminations, and more so in delayed (48 hr) retention of these discriminations, although these animals can eventually reach a criterion of accurate performance (Zola-Morgan & Squire, 1985). Furthermore, monkeys with hippocampal system damage perform more poorly when the interval between stimulus repetitions is elongated and the memory "load" increased by presenting multiple, concurrent object discrimination problems than when problems are presented one at a time with relatively short repetition intervals (Moss et al., 1981; Correll and Scoville, 1965). A similar pattern of findings is observed in rats with hippocampal system damage - they normally acquire single object discriminations but are impaired in learning the same discriminations presented concurrently (Wible et al., 1992). In addition, Staubli et al., (1984) reported that rats with entorhinal lesions acquired normally simultaneous odor discriminations presented in massed trials, but were impaired even on single discriminations if trials were spaced by 10 min. In considering the representational demands required for object discrimination, a hippocampal coding of relationships between experiences with discriminative stimuli could mediate performance, but differential responding could also be supported by hippocampal-independent acquisition of biases for individual discriminative stimuli. Thus, object discrimination performance could be supported by hippocampal-dependent or hippocampal-independent representations, and the data on object discrimination seem to fall outside of the representational distinction. Rather, it seems that impairment is observed in monkeys with hippocampal system damage only when the temporal gap that must be bridged is sufficiently long, or when the number of items that must be maintained in memory becomes sufficiently large (Correll and Scoville, 1970). As with the pattern of the findings on DNMS, it appears that the object discrimination task is primarily sensitive to the strength and persistence characteristics of hippocampal memory processing.
The data so far described are consistent with the view that a preserved short-term memory supports both DNMS performance at brief delays and single-pair object discrimination learning with brief intertrial intervals; monkeys with hippocampal system damage are impaired whenever the delay in DNMS or the repetition interval in object discrimination learning is sufficiently long. However, additional data amend this account, revealing an upper as well as a lower time- limit of hippocampal involvement. Thus, in contrast to the findings from massed training on concurrent object discrimination, monkeys with hippocampal system damage perform as well as normal subjects on a 20-pair concurrent object discrimination if the repetition rate for each problem is very long, that is, when each problem is presented only once per daily training session (Mishkin et al., 1984). This surprising finding indicates that there is an "outside" temporal boundary to memory processing mediated by the hippocampal system. According to this interpretation monkeys with hippocampal system damage perform as well as normal subjects on 24-hr concurrent discrimination because the hippocampally-mediated representation of object-memories dissipates in less than one day, making acquisition of this task independent of hippocampal system involvement in normal monkeys as well. Most relevant to the point of this review is that the representational demands of the conventional concurrent discrimination and the 24-hr delay concurrent discrimination are identical. Thus, the most parsimonious explanation of this pattern of findings is that the hippocampal system subserves a memory store of intermediate duration, at least for representations of single isolated stimuli. As will be discussed at length below, the intermediate storage of more complex information, such as spatial arrangements and other examples of relational information, involves the entire hippocampal system.
Differential contributions of hippocampal system components to intermediate-term memory for single items. The above described observations are based on comparisons of the performance of normal animals with that of animals with various subtotal or complete removals of the hippocampal system. A review of the data on DNMS in monkeys and rats with damage confined to specific hippocampal system structures indicates that the capacity for intermediate-term storage of single perceptual representations is supported by processing within one particular area of the hippocampal system. These results, summarized in Figure 2, suggest that the parahippocampal region is selectively critical to intermediate-term storage supporting recognition memory in the DNMS task. In monkeys, transection of the fornix, which disconnects the hippocampal formation from subcortical structures, does not affect performance over intermediate retention intervals, although modest impairment is observed at very long delays (Bachevalier, Parkinson, & Mishkin, 1985; Gaffan, Gaffan, & Harrison, 1984; Zola-Morgan et al., 1989b). Ablation of the hippocampal formation plus partial removal of the parahippocampal region, or cell loss restricted to the hippocampus as a result of a transient ischemic event, result in modest impairments on DNMS (Zola-Morgan et al., 1992; Zola-Morgan et al., 1989b). By contrast, selective damage to the parahippocampal region, specifically the removal of perirhinal-parahippocampal cortex (Zola-Morgan et al., 1989c; Suzuki et al., 1993) or parts of the perirhinal-parahippocampal and entorhinal areas (Murray & Mishkin, 1986; Gaffan & Murray, 1992), results in a profound deficit in DNMS performance over intermediate periods, that is, periods beyond a few seconds, with the magnitude of the impairment directly related to the length of the delay. These data clearly indicate that the cortical areas adjacent to the hippocampal formation play a more important role in intermediate-term retention of memories for specific objects than does the hippocampal formation itself. The finding that damage limited to the fornix or hippocampal formation results in a modest deficit suggests a declarative representation mediated by the hippocampal formation confers only a modest advantage in intact animals.
The findings from experiments on the effects of selective hippocampal system lesions in rats performing DNMS tasks that employ trial-unique or infrequently-used stimuli are entirely consistent with the data on monkeys (Figure 2b; see also Rawlins et al., 1991). These experiments have shown that lesions of the hippocampus or fornix produce little or no effect on recognition memory for single, non-repetitive stimuli over intermediate periods, although deficits have been observed at very long delays (Aggleton et al., 1986; Otto & Eichenbaum, 1992a; Rothblat and Kromer, 1991; Mumby, Wood, & Pinel, 1992). By contrast, damage to the perirhinal and entorhinal cortex results in the pattern of delay-dependent impairment in intermediate-term memory that parallels the phenomena in monkeys with the same cortical damage (Otto & Eichenbaum, 1992a). Rats with combined hippocampus-amygdala lesions that included some damage to the cortex, but not isolated hippocampal lesions, demonstrated a similar impairment (Aggleton et al., 1989). Jarrard (1993, p. 15) came to a similar conclusion based on his studies of selective hippocampal and entorhinal lesions in rats. He found that entorhinal lesions, but not hippocampal lesions, resulted in delay- dependent impairment in intermediate-term memory in an object recognition task and impaired acquisition of a concurrent object discrimination. Combining the data for versions of the DNMS task across species, it is clear that the parahippocampal region plays a much larger, if not fully selective role compared to that of the hippocampal formation in intermediate-term retention of representations for single sensory cues. Although a fully formed representation mediated by the hippocampal formation might confer some advantage to normal animals on DNMS over that of animals without normal hippocampal function but with an intact parahippocampal representation, the advantage is relatively small and appears only at very long retention intervals.
Combining the findings across these studies, and anticipating findings from the neurophysiological studies to be discussed below, one can approximate a set of "persistence functions" for distinct brain structures as a way of envisioning how object recognition performance is supported by different components of our model. Figure 3 provides idealized persistence functions generated from a single exposure to an object, compared for the anatomical components of the model. Neocortical areas are seen to support only a short-term trace for such memories, illustrated as a rapidly decaying retention function. The parahippocampal region is seen to add a more persistent representation, one that would support object recognition for intervals of several minutes. By contrast, hippocampal-independent systems provide a persistent (hence LTM) representation of relatively low "retention strength" after a single exposure. The actual performance of intact animals at any particular retention time can be envisioned as the sum of the strengths contributed by the neocortical, parahippocampal, and hippocampal-independent representations (excluding, for the purposes of this exposition, the contribution of hippocampal-dependent LTM). By comparison, the performance of animals with damage to the parahippocampal region can be envisioned as only the sum of the neocortical and hippocampal-independent representations (again, excluding hippocampal-dependent LTM), resulting in a steep delay-dependent performance curve. For a final comparison in this figure, an additional idealized curve is included to represent the strength of a representation following a stimulus that had been processed sufficiently by the hippocampus itself, permitting the establishment of a robust and potentially permanent memory trace. Such a hippocampal-dependent LTM representation, while likely to be engendered by significant life events, is improbable following single exposures to cues in the object-cued tasks described above. Rather, long term retention would normally be built up through repetitive exposures.
4. Representational Properties of Hippocampal-Dependent and Hippocampal-Independent Memory. Even as the early studies on H.M. and other amnesic patients generally concluded that the memory deficit associated with medial temporal damage extended across all modalities of testing materials, some exceptions were noted, particularly the acquisition of motor skills (Milner, 1962; Corkin, 1968) and perceptual "priming" tasks (Warrington and Weiskrantz, 1968). Over the last 25 years it has become increasingly clear that these "exceptions" were merely examples of a large range of preserved learning capacities that fit within the same temporal domain associated with impairment on other learning and memory paradigms (see Cohen, 1984; Cohen & Eichenbaum, 1993; Corkin, 1984; Squire, 1987, 1992). Thus, for example, amnesic subjects can learn to draw from a mirror image view and to identify pictures from partial information even when repeated presentations were widely spaced, and even when they cannot explicitly recall the learning experience. Other experiments have indicated that learning involving the identical materials may be either substantially impaired or completely spared in amnesic subjects depending on whether memory was assessed by asking subjects to employ conscious (explicit) recollection of the study phase and the materials presented therein versus assessing memory indirectly or implicitly by evaluating changes in actual processing performance caused by exposure to test materials (e.g. Graf et al., 1984).
Substantial consensus has been reached in characterizing the amnesia consequent to hippocampal system damage as a selective impairment in declarative (or explicit) memory. This kind of memory is exemplified by our recall of everyday facts and events. Such memories can be brought to conscious recollection, are typically subject to verbal reflection or other explicit forms of recall, and can be used flexibly in a variety of situations outside that of the learning experience (Cohen & Squire, 1980; Cohen, 1984; Squire, 1987). Nearly all conventional assessments of human recall and recognition require or emphasize the explicit expression of memories, and performance across this wide range of tasks is severely impaired in amnesia associated with hippocampal system dysfunction. By contrast, the domain of preserved learning in amnesia has been described as procedural memory, a collection of capacities that can be revealed in the absence of conscious recollection or verbal reflection, and are expressed implicitly through enhanced or altered performance during repetition of relevant aspects of the learning experience (Cohen & Squire, 1980; Cohen, 1984; Schacter, 1987). Most of the behavioral paradigms in which procedural memory has been studied involve the acquisition of various perceptual, motor, or cognitive skills, and several forms of adaptations and biases in performance (reviewed in Cohen, 1984; Cohen & Eichenbaum, 1993; Squire, 1992). Perhaps the most studied of these paradigms involves the "priming" phenomenon by which normal subjects are more likely to complete or recognize pictoral or verbal items from familiar partial information, or to read more quickly or name more quickly recently experienced text or pictures of objects. Human amnesics often perform normally on such tasks, although they are unable to consciously recollect their testing experiences. Notably, such enhanced performance in both normal subjects and amnesics is inflexible in that it succeeds only when the physical form of the test item closely matches that of the previously experienced item (e.g. Schacter, 1985; Tulving & Schacter, 1990).
Animal models of hippocampal-dependent memory representation. In the development of animal models of amnesia, it became clear very early on that in monkeys, rats, and rabbits, circumscribed hippocampal system damage impaired performance in some tasks but spared performance in others. These findings, and many others in both the human and animal literature, support the notion that in animals, as in humans, there are multiple memory systems that participate in permanent memory storage and that the hippocampal system is critical only to a particular type of memory or aspect of memory representation. Several proposals have been developed attempting to characterize the form of memory representation dependent upon the hippocampal system in animals, with each proposal based on a different set of behavioral data contrasting hippocampal-dependent and hippocampal-independent forms of learning.
A critical question in this area of research involves the extent to which the pattern of impaired and spared capacities in human amnesics and animals with hippocampal system damage can be captured in a single characterization. We have proposed (Eichenbaum et al., 1992a,b), as have other investigators (e.g. Squire, 1992), that the data on human amnesia and on animals with hippocampal system damage may indeed share a common fundamental basis. Extending the notion from work on human amnesia that declarative memory mediates comparisons among items in memory and permits the flexible use of memories in novel situations (Cohen, 1984), our account suggests that the hippocampal system mediates the storage of the outcomes of these comparisons in terms of critical relations among items. Such a relational representation constitutes a memory "space", an elaborate organization that permits access to memories via novel routes and supports the flexible expression of memories via pathways not previously exercised (Eichenbaum et al., 1992b). In the absence of relational representation one would still expect intact learning that can be acquired through individual representations involving the improvement of specific motor routines, adaptations of sensory-motor responses, or the biasing of responses toward specific, isolated items. The central properties of relational and individual representations can be summarized from our comprehensive review (Cohen and Eichenbaum, 1993) as follows: Relational representations...
... are created by and can be used for comparing and contrasting individual items in memory, and weaving new items into the existing organization of memories. This form of representation maintains the "compositionality" of the items, that is, the encoding of items both as perceptually distinct "objects" and as parts of larger scale "scenes" and "events" that capture the relevant relations between them.
... support the flexible use of memories by permitting access to items from various sources and by permitting the expression of memories in various, even novel, situations.
Individual representations...
... involve the "tuning" or biasing of items within separate processing modules of the brain operating in isolation. Depending on the operating characteristics of these modules, individual representations can incorporate combinations of stimulus elements, but such processing involves the fusion of stimulus elements into a single representation lacking the property of compositionality.
... can support alterations in performance during repetitions of the processing events that occurred during the original learning experience, but are inflexible in that they cannot be accessed or expressed in novel situations.
The above characterizations of relational and individual representation (greatly abbreviated here; see Cohen and Eichenbaum, 1993, for a complete presentation) are intended to apply for both humans and animals and across all kinds of learning and memory tasks. In addition, it is critical to note that, ordinarily, both hippocampal-dependent and hippocampal-independent mechanisms are available at all times to intact humans and animals and both can contribute to performance on virtually any task. One should expect to observe species and task differences to the extent that separate species are differentially inclined to utilize hippocampal-dependent representations and to the extent this form of representation is required to support different types of performance. Indeed species differences are apparent in the general observation that human amnesia seems to be quite "global" whereas the impairment in animals with hippocampal system damage appears more limited. Apparent task differences are also prominent. For example, among the most frequently reported task differences is the severe deficit in spatial learning and apparent sparing of non-spatial learning in rats with hippocampal system damage. Nevertheless, while differences in the apparent prevalence of impairments across species and tasks are to be expected, careful behavioral assessments should reveal the same fundamental character of the impairment among species and tasks.
Expanding on this point with regard to studies on animals, the literature on the effects of hippocampal system damage in animals is voluminous and has been documented extensively elsewhere, at least with regard to many of the behavioral paradigms employed in previous decades. Exhaustive tabulations and comparisons have revealed differences in the incidence with which deficits or sparing of function are observed on various categories of tasks (see O'Keefe and Nadel, 1978; Gray, 1982; Gray and McNaughton, 1982). Nevertheless, for each such tabulation we were more impressed that impairment is not consistently associated with any particular task category. For example, a consideration of 'simultaneous' versus 'successive' discrimination reveals that, in some experiments, impairment was observed only on simultaneous discrimination (see below) and in other experiments impairment was observed only on successive discrimination (e.g. Kimble, 1963; see also Gray and McNaughton, 1983, Tables 20 & 22). In addition other general variables by which tasks can be directly compared, such as task difficulty, also are not consistently associated with impairment. Thus impairments are observed on both easy (that is, rapidly acquired) problems, such as the Morris water maze, spontaneous alternation, contextual learning, and difficult (that is, slowly acquired) tasks, such as some conditional discriminations. Furthermore, no impairment is observed on some difficult tasks, such as visual pattern discrimination, and impairment may or may not be observed on variants of tasks that are equivalent in difficulty, such as variations on odor discrimination learning (examples of each of these paradigms are cited below).
Moreover, we question the utility of straightforward tabulations because such an exercise inevitably leads to a sort of psychoarithmetic by which conclusions are made according to the simple majority of data. In our view, this approach ignores the scientific rule that even single examples and counterexamples regarding particular categories of paradigms are important and must be explained (see Cohen and Eichenbaum, 1991, 1993). Conversely, we are impressed that a mixture of data is observed within virtually every learning and memory task. Our review of this literature (Cohen & Eichenbaum, 1993) has shown that that subtle variables can dramatically affect the representational demands within any behavioral paradigm, altering the relative contribution of hippocampal-dependent and hippocampal-independent processes, by encouraging or hindering relational representations and by placing more or less demand on representational flexibility. Thus, in our view, it is fruitless to simply tabulate results from single variants on any paradigm.
Adopting a different approach, the distinctive properties of relational and individual representations can be captured in animals using properly designed behavioral testing paradigms that assess two critical aspects of memory representation. First, one can compare the performance of normal subjects and those with hippocampal system damage on variants of the same paradigm that either demand or hinder the making of comparisons and representing of relationships among experiences with multiple items. Our account predicts that animals with hippocampal system damage will perform poorly when the demand for comparing and relating items is high, but perform at least as well as normal subjects when comparisons are de-emphasized or hindered. Second, one can assess representational flexibility in the same subjects by requiring them to express successfully acquired memories in novel test situations. Our account predicts that, even when successful in learning, the memory representations of animals with hippocampal system damage will be inflexible such that they can be expressed only through repetition of the original learning experience.
To illustrate these ideas, what follows is a brief review of four general learning and memory paradigms that are currently under intensive experimental scrutiny in the animal models of amnesia literature and in which sufficient evidence exists to assess our predictions: sensory discrimination learning, place learning, working memory, and conditional and contextual learning. As with every other behavioral assessment employed in this line of research, the data for each of these paradigms on the question of hippocampal involvement is mixed. As stated above, this state of affairs is fully expected from, and can be fully explained by, our view that the procedures of virtually any learning paradigm can be construed so as to encourage or hinder relational representation and to require or eliminate the need for representational flexibility. The experiments we emphasize are those that exemplify how a consideration of these properties of hippocampal-dependent learning can clarify the pattern of impaired and spared performance within each paradigm. These data, together with parallel data from studies of various task performances in human amnesia, have been reviewed more completely elsewhere (Cohen and Eichenbaum, 1993).
Sensory discrimination. The findings on discrimination learning involving specific, non-spatial stimuli have been reviewed extensively and indeed are quite mixed (e.g. O'Keefe and Nadel, 1978, Table A17; Gray and McNaughton, 1982, Tables 20 & 22). In some experiments, such as those involving the sensory discrimination tasks described above, the pattern of impaired and spared performance can be explained by differences in the memory delays or the memory load incurred when many items must be learned and remembered over the same period. However, in other studies temporal factors do not fully account for the pattern of performance. Instead the pattern of performance by animals with hippocampal system damage can be related to differences in the representational demands across task variants.
For example, in our own work, we have observed impairment, no-effect, or facilitation of learning on different variants of the same odor discriminations in rats with damage of the hippocampal system (Eichenbaum et al., 1988, 1989; Otto et al., 1991). In each variant of the task, the memory delays and load were similar, but other variables involving presentation of stimuli and response requirements were manipulated so as to alter the demand on relational representation and representational flexibility. First, we found that rats with fornix lesions were severely and persistently impaired in learning odor discriminations when they were required to choose between stimuli presented simultaneously and in close juxtaposition, thereby encouraging stimulus comparisons and selection among alternative stimulus choices. By contrast, they performed as well, and sometimes better than normal rats when the same stimuli were presented individually and successively on separate trials and no stimulus choice was required. Indeed, under conditions that strongly hindered comparisons between cues and eliminated differential choice responses, thus encouraging individual representations for each cue and biasing the execution of a single response, rats with hippocampal system damage outperformed normal rats. Second, we found that even under those conditions in which rats with hippocampal system damage succeeded in learning, their memory representations were abnormally inflexible. In one of our odor discrimination tasks, normal rats could identify familiar odors when "mispaired" in unusual configurations, but rats with hippocampal system damage performed as if presented with unfamiliar stimuli in such tests. Combined, these findings demonstrate the importance of determining the nature of the memory representation used by normal subjects, rather than focusing on the type of stimulus materials, formal description of the task, or even the magnitude of the initial learning impairment. Moreover, these results illustrate the usefulness of the two stage strategy proposed above for assessments of relational memory processing and suggest it can be used to sort out the mixture of results on many variations of this paradigm.
Spatial learning and memory. Studies on spatial learning also involve a mixture of findings (see O'Keefe and Nadel, 1978, Tables A18 & A20; Gray and McNaughton, 1982, Tables 21 & 25). Here we will summarize two experiments in which the cues guiding performance were identical across variants of a task, but other aspects of the cues or performance requirements were altered. One of these experiments involved variants of a plus-maze task where the location of a reward was guided by a set of extramaze cues (O'Keefe and Conway, 1980). In both variants the rat began trials in any of three arms of the maze and had to go a fourth, goal arm as directed by the location of the cues. In one variant of the task the cues were distributed around the room and located between maze arms, thus encouraging the rats to use both the spatial relations among the cues themselves as well as relationships between the cues and the start position to guide performance. In the other variant of the task, the cues were clustered together and located just off the end of the goal arm, thus providing the view of a single compound-cue contiguous with the reward. Rats with hippocampal system damage were impaired in learning when the cues were distributed, but performed at least as well as normal animals when the cues were clustered. These findings are consistent with our view that variations in the arrangement of the same cues so as to encourage a relational representation confers an advantage for intact subjects over the limited strategies available to animals with hippocampal system damage, whereas providing an opportunity to adopt a bias towards a single, albeit complex, cue eliminates this advantage and makes the strategies available to animals with hippocampal system damage as efficient as those that can be employed by normal subjects.
A second study that employed appropriate variations on a task involving identical spatial cues focused on spatial learning in the Morris water maze (Eichenbaum et al., 1990). In this paradigm the apparatus involved a 3-m diameter tank filled with water that was made opaque by the addition of milk powder (Morris, 1984). A small escape platform was submerged just under the surface of the water so that it could not be seen by a swimming rat. On each training trial a rat began from one of four starting positions and learned to swim to the platform guided by various distal cues surrounding the water tank. This variant of the task strongly encourages the comparison of cues seen from the multiple views of the environment, and a representation of those cues in terms of their spatial relations and their relations to different start positions. Rats with hippocampal system damage were severely impaired in learning this version of the task (Morris et al., 1982). We developed a variant of the task in which the same arrangement of spatial cues was employed but the demand for comparing views of the maze was eliminated by consistently starting the rat from the same start position, thus permitting subjects to adopt a bias in swimming towards a single, albeit complex, view in the test room. Under these conditions, rats with hippocampal system damage learned nearly as rapidly as intact rats. However, in subsequent testing where rats were required to navigate to the escape platform from new starting positions, normal rats could use their representation of spatial relations among cues in the water maze but rats with hippocampal system damage performed poorly when swimming from novel starting positions, sometimes never locating the platform in the otherwise familiar environment. Combining the results of this study and the O'Keefe and Conway experiment described above, both sets of findings show that either impaired or intact maze performance may be observed in rats with hippocampal system damage, depending on the extent to which subjects are encouraged to encode cues in terms of their spatial relationships. In addition, parallel to the findings on olfactory discrimination, these results show that, even when rats with hippocampal system damage are successful, their representation does not support flexible memory expression. Finally, our review of the data on spatial learning suggests that the disproportionate reporting of deficits within this "modality" is a consequence of the extraordinarily strong demand that spatial processing puts on relational representation and representational flexibility rather; the findings do not require a that hippocampal relational processing is unique to place learning.
Working memory tasks. Working memory, as defined by Olton and colleagues (1979), refers to memory that is useful for only a single test trial. Superficially, it might seem that the relevant memory demand in working memory, especially working memory for non- spatial stimuli, focuses on intermediate-term storage and recognition of individual stimulus items, similar to the demands of the DNMS tasks described above. However, a more thorough comparison of the representational demands of working memory and DNMS tests indicates that working memory requires a relational representation.
To understand how the working memory component of performance in the non-spatially cued radial-arm maze also requires a relational representation, it is important to consider the differences in the working memory demands of this task from the kind of memory required in recognition paradigms like the DNMS tasks discussed above. Consider for example the behavioral paradigm widely used to test working memory capacity - the radial arm maze task. The apparatus involves an elevated maze comprising a set of arms radiating from a central platform like spokes on a wheel. On each trial all of the arms are baited and the rat is allowed to retrieve the rewards. The most efficient foraging strategy involves remembering which arms have been entered and not revisiting those arms within a trial. The cues available that guide performance are either spatial cues provided by the extra-maze environment or visual and/or tactile cues that distinguish the maze arms. Notably, in either variant of the task, all the cues are typically available throughout performance within each trial; the rat is not prevented from seeing the cues between making its delayed choices. Consequently, remembering the stimuli that identify each arm is not required. Rather, what must be remembered during a trial are the behavioral events associated with entering each arm, that is, rats must remember their previous episodes of entering arms in addition to, or instead of, remembering the distinguishing qualities of the arms themselves. By contrast, recognition memory in tasks using trial-unique or infrequently-presented single cues can be accomplished by a direct perceptual match between test stimuli and a memory representation of the sample stimulus, that is, by remembering and recognizing the stimulus itself without necessarily remembering the episode of interacting with the sample.
Thus, in our view, working memory involves the encoding of specific "episodes" associated with perceptually defined maze arms, and not a perceptual trace of the arm itself (cf. Olton, 1984). The events that characterize each episode in working memory are defined in terms of how they compare and contrast with other experiences with the same stimuli - as such, episodic memories clearly exhibit the properties of relational representation. This is not to say that it would be impossible to guide performance in working memory tasks on the relative familiarity of cues - presumably arms with distinctive cues visited within a session would have stronger relative familiarity than arms not visited in that session. Individual representations based on the enduring strength of decaying sensory representations might be used to support performance even when memory for the previous episodes within a session is unavailable. Consistent with this possibility, rats with hippocampal damage are only modestly impaired, or impaired only on some variants of the non-spatial radial arm maze task (Winocur, 1982; Nadel and MacDonald, 1980; Jarrard, 1986).
Notably, the pattern of performance on several other spatial and non-spatial tasks are explained by a strong working memory component. For example Thomas and Gash (1988) employed a T- maze task that involved a combination of a consistent reward association (always go right at the first choice point) and a working memory component (turn in the direction opposite to that last taken at the second choice point). Hippocampal system lesions had no effect on the consistent component but resulted in severe impairment on the working memory component. This finding is reminiscent of the frequent (but not reliable) finding that rats with hippocampal damage normally learn simple spatial discriminations in a T- or Y-maze but are impaired at spontaneous or learned alternation guided by the same available cues (O'Keefe and Nadel, 1978 - compare Tables A15, A16 & A25). Rats with hippocampal system damage are also impaired on performance in a go, no-go alternation task (Winocur, 1985, 1991; see also Jarrard, 1975) and in delayed matching to sample guided by trial- repetitive brightness cues (Winocur, 1992; Rafaelle and Olton, 1988). Importantly, severe impairments are also observed on other tasks that do not require working memory (e.g. the conventional version of the Morris water maze), indicating that the demand for working memory is not a prerequisite for hippocampal involvement. However, as these findings show, hippocampal system function is required for normal performance in both spatial and non-spatial tasks where there exists a strong demand for working memory.
Conditional and contextual learning. A large number of studies have employed variations of conditional and contextual learning tasks, and the use of these tasks has come under considerable scrutiny only in recent years. However, as will be shown, reports on the effects of hippocampal system lesions on this collection of paradigms are mixed, as was the case with the other paradigms described above (e.g. see Jarrard, 1993). Nevertheless, in our view, these data can be reconciled by a consideration of the demand for relational representation present in variants of the tasks employed.
Conditional learning tasks are like sensory discriminations except that the reward assignments for particular cues are dependent on the presence of another cue. In some cases the conditional cue is itself a specific stimulus appearing concurrent with or prior to a punctate "explicit" cue that is presented just prior to a reinforcer; in other cases the conditional cue involves "background" stimuli (i.e. the context) present throughout a particular training session. In all such tasks explicit cues are presented both with and without a conditional cue, or their combination constitutes the conditional cue, and the multiple stimulus situations must be differentiated in accordance with the conditional cue(s). One might presume that conditional and contextual learning paradigms provide particularly good tests of our account because these types of learning tasks would seem to require the encoding of relationships between explicit cues and other conditional cues or a context as could be mediated through hippocampal-dependent mechanisms. However, there are other ways in which conditional and contextual cues can be encoded that do not require relational representation. In some situations conditional cues can act as general "facilitators" that signify and increased probability of reinforcement for any following stimulus. The mechanisms for such facilitation would be expected to involve the acquisition of hippocampal-independent biasing of responses. In other situations, conditional cues can be fused with the explicit cue to form a compound or "configural" stimulus. To the extent that explicit cues can be encoded along with conditional or contextual cues as distinguishable compounds, learning guided by unique configural cues could also be supported by individual representations through hippocampal-independent mechanisms for recognition and acquisition of biases, just as discussed above with regard to sensory discrimination. Assuming that subtle task parameters might alter the extent to which subjects encode the cues in terms of relationships between perceptually independent stimuli versus general facilitators or unique configural stimuli, one would expect the literature on conditional and contextual learning to be mixed with regard to the effects of hippocampal lesions. Such is indeed the case: Different investigators, using different variations of these tasks, have found contradictory results and reached opposite conclusions.
One might conclude that this state of affairs renders studies on conditional and contextual learning useless with regard to the goals of this review. However, there are ways to characterize the nature of representation employed in conditional and contextual learning, specifically by tests for general facilitation (see below) and for the property of compositionality described above in distinguishing relational and individual representations. If indeed the representation used during learning is relational, subjects should still be able to recognize and respond appropriately to the individual cue elements in accordance with the specific acquired relationships. Conversely, if the representation is configural, subjects would not be expected to recognize stimulus elements in isolation. Some of the paradigms used in conditional and contextual learning preclude such an analysis, such as when all presented combinations of cue elements are differentially reinforced during original training. However, in other paradigms probe tests directed at the nature of representation have been performed, and have shown that animals with hippocampal system damage sometimes abnormally employ a configural strategy, as will be seen below, consistent with our view that configural representations occur independent of hippocampal function. These findings suggest that animals with hippocampal system damage actually resort to configural representations in the absence of the capacity for relational processing.
The distinctions between "relational" and "configural" representations and "facilitators" in conditional and contextual learning can best be illustrated by considering examples of several variants on these paradigms. Below we will discuss examples of both "conditional learning" tasks, where subjects are required to employ the conditional cues to distinguish ambiguous explicit cues, and "contextual learning" tasks, which involve a set of static background cues that are not varied during training and thus are only incidental to the explicit cue and it consequences (such cues are typically altered only in a post-training test phase).
Conditional learning tasks. Three different general variants of conditional learning have been reported. In one of these, called Pavlovian feature-positive conditioning, an explicit cue is followed by reinforcement only if it is preceded or occurs concurrent with another, conditional cue. For example, in a study by Ross et al., (1984) presentation of a tone was followed by delivery of a food reward only if this cue was preceded by a conditional light cue; thus the light was a conditional cue for the efficacy of the tone. Rats with hippocampal system damage were severely impaired in acquisition of the conditional response, but acquired normally appropriate discriminative responses to other simple, i.e., non-conditional stimuli (a clicker and a noise). Jarrard and Davidson (1990, 1991; Davidson and Jarrard, 1989) confirmed these results when the lesion involved aspirations of the hippocampus, but not with more selective neurotoxic lesions. Furthermore, they showed that the conditional cue acted as a general facilitator of responding to any cue of ambiguous significance, rather than as an element of either a specific relational or configural representation with the explicit cue, at least in their version of the task. Thus this paradigm probably does not serve to test the relational (or configural - see below) hypothesis. A more valid test may have been provided in a study by Gallagher and Holland (1992) who employed a variant of this task in which two auditory cue elements served as CS+ and CS- cues and a panel light conditional cue presented along with the explicit cue served to reverse the valence of each on some trials; the task thus included both "feature-positive" and "feature-negative" properties of the conditional cue, making it unlikely that the conditional cue could serve as a general facilitator or inhibitor. Indeed, in separate experiments, Holland (1991) has shown that such training does involve specific configural representations of the conditional and explicit cues. Notably, Gallagher and Holland failed to replicate the Ross et al. (1984) result using this paradigm, although comparison of the results across these studies is obscured by the fact that Gallagher and Holland also used the more selective type of hippocampal system lesion. These findings are consistent with our proposal that subtle alterations in task procedures can dramatically change the nature of the memory representation used from one that demands hippocampal function to another that does not. Further evidence, more directly supporting this notion, is the finding that whether or not animals with hippocampal lesions were impaired on feature positive conditioning in an eyeblink conditioning task depended on which of two cues was the conditional cue and which the explicit cue (Loechner and Weitz, 1987). In addition, for at least some variants of this paradigm, selective neurotoxic lesions of parts of the hippocampal formation produce weaker effects than more complete hippocampal formation lesions produced by other methods. The issue of complete versus partial hippocampal formation lesions (see p. 6) has been especially problematic for interpreting discrepancies in the findings on conditional learning. Of particular value would be studies that examine the effects of selective neurotoxic lesions of the entire hippocampal formation, that is, the hippocampus, dentate gyrus, and subiculum.
A second variant on conditional learning is the negative patterning paradigm (Sutherland and Rudy, 1989). In this task there is no separate conditional cue; rather, subjects are trained to distinguish the separate presentation of two cue elements (e.g. a light and a tone), each of which is associated with reinforcement for an operant response, from the simultaneous presentation of the combined cues, which predicts non-reinforcement. In their original study using this task, Rudy and Sutherland (1989) reported that rats with hippocampal system damage could not acquire the negative patterning task, even though they performed normally on a simple discrimination between the light and a different tone. However, Davidson et al., (1993) failed to replicate this result, regardless of whether the lesion method involved their highly selective neurotoxic technique or the same less-selective lesion technique of Rudy and Sutherland (1989). The failure to replicate may be attributable to procedural differences between the tasks (e.g. specific differences in the response requirement) that resulted in different representational strategies by the subjects. Overall, the mixture of results across studies on feature-positive/negative and negative patterning tasks serves to highlight the need for closer examination of specific training procedures as well as lesion methods. Nevertheless, in both the Davidson et al. (1993) and Gallagher and Holland (1992) studies, it seems likely that configural representation, rather than general facilitation/inhibition, was employed by rats, suggesting the hippocampus itself may not be required for this form of representation thereby explaining the failure of these investigators to find hippocampal-lesion effects (see below).
A third variant on conditional learning will be termed here "conditional discrimination". In several variations on this task, multiple conditional cues serve to qualify the reward associations of subsequent explicit cues. In an early variant of this task (Kimble, 1963) the color of a Y-maze (white or black) qualified the placement of a reward in the left or right arm of the maze. In other experiments (Hirsh, 1974; Hsaio and Isaccson, 1971), the motivational state of the animal (hunger or thirst) predicted the placement of reward in a T-maze. In more recent variants on this task, Sutherland et al. (1989) used the illumination condition of a start box in a Y-maze to qualify whether a reward would be located in a white or black goal arm, and Whishaw and Tomie (1991) used the odor of a string to qualify whether pulling a thick or thin string produced a reward. In each of these tasks animals with hippocampal system damage were impaired in learning the conditional discrimination while, at least in the more recent studies, the same animals demonstrated intact learning of simple discriminations guided by the same cues.
In other studies on spatial and non-spatial conditional discriminations the effects of hippocampal system damage were also mixed (e.g. Markowska et al., 1989). Most relevant to the views expressed here was the finding that monkeys with hippocampal system damage were impaired when required to use the location of explicit cues within a visual scene as the conditional cue, but unimpaired when required to use different places in the test room composing completely different background scenes as the conditional cue (Gaffan and Harrison, 1989; Murray et al., 1989). Combined with the findings from other studies of conditional discrimination described above, the results strongly implicate hippocampal system function in non-spatial conditional discrimination learning, although too little work has been done to characterize the form of representation employed in any of these paradigms.
Contextual learning tasks. In studies on contextual learning there is also a mixture of findings of sparing or impairment. Furthermore, across variants on this paradigm hippocampal system damage can result in either impaired or abnormally strong utilization of contextual cues. Some studies employ paradigms where the retention of a conditioned response is dependent on context, such that in intact animals good retention is observed when they are tested in the same context used during initial learning and poor retention is observed when the context is changed. For example, Penick and Solomon (1991) found that classically conditioned eyelid responses are context dependent in normal rabbits, and Good and Honey (1991) described various forms of appetitive conditioning that are similarly context-dependent in normal rats. In both studies animals with hippocampal system damage failed to demonstrate context dependency. In addition, during conditioning of fearful responses to an explicit cue (typically a brief tone) there is an incidental conditioning to the contextual stimuli, observed as increased fearful behaviors when the subject is re-exposed to training context without or prior to another presentation of the explicit cue. Hippocampal system damage results in poor conditioning to the context in this paradigm, even though conditioning to the explicit cue, even when presented in a different context, is spared (Kim and Fanselow, 1992; Phillips and LeDoux, 1992; Selden et al., 1992). Importantly, rats with hippocampal system damage condition normally to the context when the explicit cue is omitted during training, indicating that processing of the same contextual cues is possible even in the absence of hippocampal function (Phillips and LeDoux, 1992). This finding clearly demonstrates that the hippocampal system is not required for processing either the explicit cue or the context, but is critical for normal representation of the relationship between these cues.
The story on contextual learning becomes more complicated when one considers other findings indicating that, in other training paradigms, animals with hippocampal system damage show abnormally strong dependence on contextual cues. Thus, Winocur and Olds (1978) reported that rats with hippocampal system damage showed abnormally poor retention of a pattern discrimination when tested in a different context, and conversely, abnormally good performance in reversal of the discrimination in a different context. Winocur and colleagues (1987) also reported that while normal rats show an aversion to an environment proportional to the probability of a CS predicting shock, rats with hippocampal system damage conditioned strongly to the environment at all levels of shock predictability.
The mixture of results across these variants of contextual learning paradigms makes it clear that the explicit cues and contexts can be related to one another using different forms of representation, leading to opposite effects of hippocampal system damage. It appears that, in situations where the explicit cue is particularly salient with respect to the context, animals with hippocampal system damage condition extraordinarily well to these cues, a phenomenon reminiscent of enhanced conditioning of individual associations in some discrimination paradigms (see above). To the extent this occurs, the observed context-dependence in animals with hippocampal system damage might be reduced. Conversely, in situations where the explicit cues are not strongly salient over contextual cues, animals with hippocampal system damage might be abnormally inclined to encode both types of cues as a configural cue, making performance overly dependent on presentation of their combination during testing. An abnormal tendency to encode cues this way has been observed when discriminative cues were closely juxtaposed in monkeys (Saunders and Weiskrantz, 1989) and in rats (Eichenbaum et al., 1989). In both these studies, animals with hippocampal system damage, unlike intact animals, failed to recognize familiar discriminative stimuli when they were presented in novel combinations; that intact animals could handle the familiar stimuli in novel combinations indicated they had stored relational representations of the perceptually independent cues, with compositionality, and could use the relational representations flexibly to make novel judgements. These results, combined with observations of intact configural association in some of the conditional learning tasks described above, supports our suggestion that the acquisition of configural representations is accomplished outside the hippocampal system. Conversely, the combined findings support the notion that the hippocampal system is critically involved in representation of perceptually independent stimuli in terms of their relevant relationships.
Summing up about paradigms employed in assessments of the representational distinction. Despite the many differences in procedures used between and within the various paradigms and despite the absence of conclusive data on the nature of the representation employed by normal animals in many of these tasks, some general conclusions can be drawn from the neuropsychological studies (see Cohen & Eichenbaum, 1993, for a more complete treatment). First, impairment or intact performance may be observed on virtually any learning paradigm depending on task parameters that encourage comparing and contrasting items and representation of the relevant relations between them. Second, even when successful in learning, animals with hippocampal system damage are severely restricted in the flexibility of their expression of learned performance. The demand for flexibility is implicit during training in some tasks, for example in the conventional version of the Morris maze task where the starting position is varied across trials, and in working memory tasks where the significance of familiar cues is constantly changing. In other tasks, such as discrimination learning and some conditional and contextual learning paradigms, learning often proceeds normally and the inflexibility of hippocampal-independent representations is revealed only in tests that probe the nature of the successfully acquired memory. Nevertheless, examples of spared learning with representational inflexibility can be observed across all paradigms.
Critical contribution of all hippocampal system components to relational representation. A major prediction of our two-stage model of hippocampal system processing is that the hippocampal formation itself becomes critical to performance whenever the task demands require relational processing; however, our model predicts that damage to the cortical areas surrounding the hippocampus will also produce deficits on these tasks, because hippocampal processing depends upon associations with neocortex that are mediated through these areas. Consistent with these predictions, lesions limited to the hippocampal formation or fornix result in severe impairments in monkeys on tasks involving spatial cues (Mahut, 1972), in conditional discriminations involving the configurations of scenes (Murray et al., 1989; Gaffan & Harrison, 1989), or the flexible expression of object associations (Saunders & Weiskrantz, 1989). The data from a large number of experiments on rats parallel these findings. Selective hippocampal ablation, fornix transection, or damage to the parahippocampal region have all been reported to result in severe impairment on performance in learning and memory tasks that involve sensory discrimination, spatial learning, working memory, and conditional or contextual learning (see reviews by O'Keefe and Nadel, 1978; Gray and McNaughton, 1982; Olton et al., 1979), and on the flexible expression of odor- or place-guided memory (see Eichenbaum, 1992; Eichenbaum et al., 1992b).
In experiments directly comparing the effects of fornix lesions and/or removal of the complete hippocampal formation with that of parahippocampal damage, equivalent deficits were observed on measures of spatial working memory (Jarrard, 1986; Olton et al., 1978, 1979, 1982). In addition, non-spatial working memory can be impaired after damage limited to the hippocampal formation, as well as after damage to the parahippocampal region (Olton and Feustle, 1981; Jarrard, 1986; Winocur, 1982; Nadel and MacDonald, 1980). Similarly, selective damage to the hippocampal formation or fornix impairs performance on DNMS tasks involving the same two spatial (Aggleton et al., 1986) or non-spatial (Rafaelle & Olton, 1988) cues on each trial. Even though these paradigms use testing procedures similar to those employed in tests of recognition memory, they emphasize working memory rather than recognition memory because the same cues are employed repetitively on each trial, minimizing the usefulness of perceptual familiarity as a cue (cf. Rawlins, 1991). Thus in both monkeys and rats, the data consistently indicate that the entire hippocampal system, that is, the hippocampal formation, its subcortical connections through the fornix, and the adjacent parahippocampal region, is required to support relational representation.
Intermediate-Term Storage of Relational Memories. Earlier in this review we distinguished a selective role for the parahippocampal region in the intermediate-term storage of individual representations from the involvement of the hippocampal formation in relational memory processing. This delineation leaves open the question of whether the parahippocampal region itself can support intermediate- term retention of relational as well as individual representations in the absence of normal hippocampal function or whether an intact hippocampal formation is required for this aspect of memory as well. The answer seems to be that processing within the hippocampal formation plays a critical role in the establishment of relational representations even for intermediate retention periods. Thus, intermediate-term recognition memory for places, or relations between particular objects and places, is impaired after damage to the hippocampal formation or fornix just as it is after damage to the parahippocampal region. For example, selective ablation of the hippocampal formation results in an impairment on delayed recognition of the place where an object was seen previously (Parkinson et al., 1988). Transection of the fornix results in a selective impairment on delayed spatial and conditional spatial tasks (e.g. Mahut, 1972; Murray et al., 1989). Selective hippocampal lesions also result in a severe deficit in rats performing delayed match-to- place tasks in the water maze and Y-maze, and in spatial non- matching and memory for the order of arms visited on the radial maze (Kesner & Novak, 1982; Morris et al, 1990; Olton et al., 1979; Olton, 1986). Selective hippocampal lesions also result in delay-dependent impairment in intermediate-term memory for non-spatial information in delayed go/no-go alternation (Jarrard, 1975; Winocur, 1985, 1991), non-spatial matching and non-matching to sample with trial-repetitive cues (Rafaelle and Olton 1988; Winocur, 1992), and in delayed responding on a conditional discrimination (Winocur, 1991). In many of these tasks, performance that requires only short-term memory has been shown to be spared, indicating that hippocampal function is not required for relational judgements per se, but rather for memory based on such judgements.
In addition, there are other behavioral paradigms in which the critical cues would superficially seem not to require relational representation, yet for which damage to the hippocampal formation results in a severe deficit in intermediate-term retention. We will consider two such paradigms here: timing and trace classical conditioning. For each of these, the relevant behavioral tasks were ostensibly created to test memory capacity for single stimulus events. However, we will argue that a careful analysis indicates that the critical memory processes supporting performance in these tasks likely requires relational processing rather than, or in addition to, storage of single perceptual representations, thereby bringing these data into line with the model presented here.
Timing. Rats with damage to the hippocampal formation or fornix are impaired on tasks that involve a requirement for responding at a specific minimum interval following a previous response, that is, timing. In one task that is often used to study the neural substrates of timing, rats receive differential reinforcement for low response rates (DRL) of an operant behavior (typically a bar press). This task requires simply that a rat wait for a fixed period between each behavioral response in order to receive a reward; premature responses reset the reinforcement schedule. Rats with hippocampal system lesions are severely impaired on this task if the waiting period is long enough and the response time is unsignalled (e.g. Clark and Isaacson, 1965; Sinden et al., 1986; Braggio and Ellen, 1976; Boitano et al., 1980); the deficit is apparent whether or not rats are allowed to participate in "collateral" behaviors that might exploit hippocampally- dependent, spatially defined behaviors to bridge the waiting period (Rawlins et al., 1983). These findings, combined with other data indicating that rats with hippocampal system damage demonstrate a foreshortened timing function when rewarded for responses at a fixed minimum interval (Meck et al, 1984; Olton et al., 1987; Meck, 1988), indicate that the hippocampal formation is important for accomplishing internally based timing.
No studies have compared the effects of damage to the parahippocampal region versus the hippocampal formation on DRL or other timing tasks. Nevertheless, to the extent that timing requires only the representation of an isolated stimulus event, the severity of the impairment after selective hippocampal formation or fornix lesions would seem to call into question our contention that the hippocampal formation is not critical to memory for single isolated cues. However, one important consideration involves the nature of the event that is remembered over the timed interval and how this memory supports timing. Our view is that the remembered event in timing tasks likely involves a relational rather than an individual representation. While it is not clear what event is being remembered in DRL, note that all the perceptual stimuli that could be used are available throughout the memory delay period. Thus, as was the case with working memory tasks, it is the behaviorally defined "episode" with such stimuli and not the memory of any particular stimulus item that apparently drives performance. As concluded above for working memory tasks, such episodic memories likely depend on a relational representation, and therefore on the integrity of the hippocampal system (see Cohen & Eichenbaum, 1993).
Trace conditioning. Another finding that is superficially incongruous with our model is the observation that damage to the hippocampal formation disrupts "trace" but not "delay" classical conditioning of the eyeblink response in rabbits (Moyer et al., 1990). To understand how the hippocampal formation itself and the kind of representation processed in the hippocampus might contribute to trace classical conditioning requires a consideration of the processing requirements and anatomical pathways supporting different forms of classical eyelid conditioning. In the standard delay version of the conditioned eyeblink task a tone CS precedes and overlaps with a UCS that evokes a reflexive eyeblink. Under these conditions learning to blink during presentation of the CS and prior to UCS onset occurs gradually, and hippocampal damage does not affect learning, consistent with the view that the hippocampal system is not critical to the acquisition of learned responses to individual stimuli (see above). In contrast, if the duration of the CS onset is abbreviated so as to create a 500 msec gap between CS offset and UCS onset (trace conditioning), hippocampal lesions result in a severe and lasting conditioning impairment (Solomon et al., 1986; Moyer et al., 1990).
A consideration of the neuroanatomy of hippocampal- brainstem connections provides an insight into how information acquired by the hippocampus could contribute to this form of conditioning. There are now converging data indicating that the pathways supporting classical eyelid conditioning are localized in the brainstem, and that acquisition is subserved by neural plasticity within the cerebellum (Thompson, 1986; Krupa et al., 1993). However, unit recording studies have demonstrated that the hippocampus receives and processes information about stimulus events associated with eyeblink conditioning in both the delay (Berger et al., 1976) and trace (Solomon et al., 1986) versions of the task. The hippocampal representation is not a simple stimulus-elicited activity but, in the early trials of trace conditioning, involves the activation of some form of neural representation at the offset of the CS that persists throughout the trace interval (Solomon et. al., 1986). Thus the hippocampus forms and maintains a representation of the stimulus events in this task even though a specifically relational representation is not required in this type of learning. Importantly, this representation can reach the cerebellar circuit via multisynaptic projections involving subicular-mammillary and septal-habenular pathways to the ventral pontine and cerebellar nuclei (Berger et al., 1986). In this way the hippocampal representation of the CS is available to the cerebellar circuit, and its lasting nature could serve in the place of a CS for a cerebellar associative mechanism that requires CS-UCS contiguity. When CS-UCS contiguity is broken, the hippocampal system contributes a CS-elicited representation that persists for the trace interval and is thus contiguous with the UCS. When this representation is prevented by a hippocampal lesion no other contiguous stimulus is available and conditioning is prevented. We assume that neural activity in the hippocampus reflects the representation of particular stimulus episodes even though a relational form of representation per se is not required and makes no special contribution to trace or delay classical conditioning. Nevertheless, the persistence of this representation can be exploited to satisfy the contiguity requirement imposed by CS-US gap in the trace conditioning paradigm.
Collectively, the observations on place recognition, spatial and non-spatial working memory, and timing are consistent with our hypothesis that hippocampal processing makes a significant contribution to memory processes in a variety of circumstances in which intermediate-term retention depends on representation of relations among cues (e.g. spatial memory and object relations) or on specific episodes with familiar stimuli (working memory and timing). In addition, the data on classical conditioning suggest a mechanism by which a persistent hippocampal representation may be exploited to bridge a brief temporal gap in the absence of explicit CS-UCS temporal contiguity. Conversely, the circumstances in which the parahippocampal region can support memory in the absence of normal hippocampal function are limited to tasks in which current stimuli can be matched to an iconic (i.e. perceptually-based) representation that can be used to make a perceptual familiarity judgement.
5. Neural Coding in the Neocortex and Hippocampal System During Memory Performance
Our model of the sequential stages of hippocampal processing is further supported by observations on the behavioral physiology of single neurons in both the parahippocampal areas and the hippocampus itself, as well as in the neocortical areas that provide the critical information upon which the hippocampal system operates. The relevant data must be derived across several studies examining the functional correlates of cortical and hippocampal neuronal activity in either rats or monkeys performing various memory tasks involving different stimulus modalities. In addition, there is a lack of clarity with regard to the site of recordings in the transitional areas between visual-temporal and perirhinal-parahippocampal cortex, making our designations of the boundaries of areas with particular coding properties tentative (see below). Nevertheless, to the extent that the results of these studies can be combined, the findings consistently indicate that there are three distinct patterns of physiological activity in single neurons that can be related to memory processing, and that each of these patterns and associated characteristics of neural coding are consistent with the processing functions assigned to specific stages in the model proposed here. These data indicate that the neocortex can support very short-term storage of specific information, that the parahippocampal area supports an intermediate-term store, and that the hippocampal formation does not store specific sensory information but rather supports relational processing.
Labile neocortical representations of individual items and events. With regard to neocortical association areas, a number of studies have revealed two different forms of memory representation by which neocortical cells sustain specific stimulus or motor encodings over short delay periods in recognition memory tasks (e.g. Mikami and Kubota, 1980; Niki and Watanabe, 1976; Fuster and Alexander, 1971; Fuster and Jervey, 1981; Fuster, 1990; Miyashita and Chang, 1988; Goldman-Rakic et al., 1990; Sakurai, 1990b). One type of memory correlate might reflect an "active" representation of specific sensory events in the form of evoked neural responses that are sustained during the memory delay period (Figure 4a, top). For example, in monkeys performing visually cued DNMS tasks, Fuster and Jervey (1981; also Fuster, 1990) and Miyashita and Chang (1988) described cells in the inferotemporal cortex that responded to the onset of sample stimuli and persisted in firing at elevated levels throughout a memory delay period during which the subject had to retain information about the stimulus in order to perform a subsequent matching response. Both stimulus-driven and delay activity in many of these cells were selective to a particular visual pattern or stimulus characteristic. In addition, several experimenters recording from monkeys performing delayed alternation and delayed response tasks have observed cells in the prefrontal area that fire at the onset of sample cues and throughout the delay period with stimulus- or behavior-selective activation. For example, Goldman-Rakic and colleagues (1990) described prefrontal cells that persisted in firing during the delay interval, encoding the position of either a particular sample stimulus or of a particular intended behavioral response that was withheld during the delay. In rats performing a tone-cued DNMS task, Sakurai (1990b) described auditory cortex cells that, like inferotemporal and prefrontal cells in monkeys, demonstrate stimulus- selective responses that persist throughout a memory delay period.
The other type of memory correlate described in neocortex might reflect a "passive" memory representation characterized by a strikingly reduced responsivity to familiar stimuli. For example, Baylis and Rolls (1987, Figure 4a, bottom; see also, Rolls et al., 1989a) described visually responsive cells in inferotemporal cortex that fired much less to a stimulus item on its immediate repetition in a serial recognition task. Furthermore, the reduction in responsiveness was, in some cells, greatest for the neuron's optimal stimulus as compared to other items, demonstrating the same item-specificity that characterizes "active" neocortical memory correlates (Baylis and Rolls, 1987; Miller et al., 1991a; 1993). Superficially, the concept of reduced activation serving as the substrate for memory storage seems counter- intuitive. Indeed, the reduction in responsiveness of inferotemporal cells to repeated stimuli has been characterized simply as an extraordinarily rapid form of habituation (Miller et al., 1991b). However, PET data from human subjects observing visually presented words (Squire et al., 1992) suggest a different interpretation. Squire et al. found that metabolic activation in an area of human visual cortex is diminished when subjects view familiar material during the retention phase of a priming task. In their interpretation of these data, they argued that diminished activation to familiar material may reflect a reduction in the neural processing or number of neural computations required for the re-identification of items that have recently been processed. Extrapolating to the level of individual neurons, the same account could explain the reduced "responsiveness" of neocortical cells to familiar stimuli - the decrement in stimulus- elicited firing reflects a diminution of the circuit activity required for stimulus re-identification. Thus "passive" memory correlates might reflect the phenomenon of "priming" in the neural circuits that process perceptual information.
Both types of neocortical memory representations can be quite labile. What we have called "active" cellular responses do not outlast the trial in which the information must be retained (e.g. Fuster, 1981). Similarly, in some areas and/or tasks, the above described "passive" correlates cannot be sustained across substantial perceptual interference. Thus, Baylis and Rolls (1987; see also Rolls et al., 1989a) reported that the reduction in responsiveness of inferotemporal neurons that normally occurred after a single exposure to a particular stimulus was abolished by presentation of a different item. Some qualifications about the lability of "passive" responses and a possible distinction with "active" responses were recently revealed in a study using a delayed match to sample task where the sample stimulus is followed by a series of choice cues. In this paradigm, "active" memory responses were abolished by presentation of another stimulus, but "passive" memory responses were observed to persist through the presentation of intervening non-matching choice cues presented within the same trial (Miller et al., 1993). In another study, active visual responses in cortical area V4 were observed to withstand interference from the presentation of other stimuli when the animal was required to remember a sample pattern while examining a series of test patterns within each trial (Maunsell et al., 1991). Notably, both the "passive" and "active" memory responses observed in these studies disappeared between trials. In the Miller et al. study, this effect was not explained by the mere passage of time, indicating that the responsiveness of visual cortical cells is "reset" when the information is no longer relevant. By contrast, in the Baylis and Rolls study, response decrements were abolished by a single intervening stimulus in a serial (running) recognition task where the monkey had to remember stimuli across a variable number of trials. The discrepancy between these findings might be related to relatively subtle differences in the visual stimuli or training procedures, but this pattern of findings might also reflect differences in locus of the recordings. Baylis and Rolls recorded primarily from cells in lateral inferotemporal cortex (in or near the superior temporal sulcus) whereas Miller et al. recorded from the anterior inferotemporal area adjacent to the parahippocampal region. Furthermore, despite differences in the short term lability of responses to single stimulus presentations, many cells in both parts of inferotemporal cortex exhibit a gradual cumulative decrement in responsiveness to items repeated many times across a testing session, suggesting that multiple experiences with a stimulus can modify the representation for longer durations (see also Rolls et al., 1989a; Riches et al., 1991). In sum, the coding properties of neocortical association cortex could support a short-term memory for single stimulus presentations using both "active" and "passive" item-specific representations but, particularly for areas distant from the parahippocampal region, these representations are short lasting and susceptible to interference; only with extended experience can these areas support recognition over delays of longer than several seconds.
Intermediate-term storage of specific events in the parahippocampal region. Unlike the neocortex, cells in the parahippocampal region do not display "active", sustained responses over memory delays. However, like neocortical association areas, parahippocampal neurons demonstrate "passive" memory representations in the form of stronger stimulus-elicited responses to novel versus familiar stimuli. Furthermore "passive" cellular memory representations in perirhinal and entorhinal cortex, like those in inferotemporal cortex, are often stimulus specific. For example, Riches et al., (1991) observed these properties in cellular activity recorded from the entorhinal cortex and, to a lesser extent, from the inferotemporal cortex of monkeys performing a DNMS task. They described cells that demonstrated stimulus-selective visual responses, and many of these cells responded more to a visual cue when it was the sample than when it immediately reappeared as the match stimulus. Similarly Sakurai (1990), recording from entorhinal cortex in rats performing a tone-cued DNMS task, described cells with stimulus-specific, sensory-elicited responses. In neither of these studies was delay-related activity in the parahippocampal region observed.
These studies show that, in contrast to most neocortical association areas, "passive" representations in the parahippocampal and immediately adjacent anterior inferotemporal region (AIT) might be relatively insensitive to interference resulting from the presentation of intervening items and trials. Brown and colleagues (1987; Riches et al., 1991) reported that neurons in parahippocampal region and AIT continued to exhibit reduced responsiveness to familiar stimuli even when other stimuli were presented during the period between an initial stimulus presentation and its repetition (Figure 4b). Again there are some inconsistencies in the data across experiments - thus, unlike Miller et al. (1991a; 1993), Riches et al did not observe a "resetting" of responsiveness between discrete trials in AIT cells. Possibly the discrepancy is due to differences in the visual stimuli or testing procedures. However, it is also possible that the discrepancy can be related to our poor understanding of the functional boundaries between visual-perceptual processing areas and the parahippocampal region. Here we will not take a firm position on whether there is a discrete functional boundary and, consistent with the view favored by Miller et al. (1993, p.1474), conclude simply that the closer to the perirhinal region the longer lasting the passive representation. To the extent this proves to be the case in more extensive comparisons, "passive" memory representations in the parahippocampal region (and in adjacent cortical areas), unlike those in more distant areas of neocortex, persist beyond the period of short-term memory and can withstand significant interference. Since these representations encode specific stimulus characteristics, they could be used to mediate a perceptual matching of characteristics of previous and current stimuli. This conclusion is consistent with our proposal that the parahippocampal region supports intermediate-term recognition of specific items in DNMS and in object discrimination learning - tasks for which performance is severely impaired after parahippocampal damage.
Hippocampal processing of comparisons among events. Studies on both monkeys (Riches et al., 1991) and rats (Sakurai, 1990) indicate that, in contrast to both neocortex and parahippocampal areas, hippocampal cells do not "actively" fire throughout memory delays in response to previously-presented sample stimuli, nor do they demonstrate "passive" representations in the form of diminished stimulus-specific responses. Some studies have reported delay-related hippocampal cellular activity, but the pattern of their activation is not consistent with the maintenance of an active perceptual representation (Otto & Eichenbaum, 1992b; Watanabe & Niki, 1985; Rolls et al., 1989a); such activity, when it does occur, likely reflects unidentified behavioral events that occur during the delay period. For example, Watanabe and Niki (1985) described cells that fired consistently at specific periods during a portion of the memory delay, but no cells demonstrated stimulus-elicited responses that were sustained throughout the delay, as should be expected of an actively maintained stimulus representation. Furthermore, Riches et al., (1991; see Brown et al, 1987) highlighted the contrast between the stimulus-specific response decrements observed for visual stimuli in the parahippocampal and inferotemporal regions with the absence of such responses in any hippocampal area. A recent report by Rolls and colleagues (1993) confirms and extends this finding. They found that hippocampal cells showed diminished responses to repeated visual cues presented with varying numbers of intervening items in a serial recognition task; these responses persisted across several intervening items but were observed for all familiar stimuli. In one study it was reported that hippocampal cells maintained "passive" memory representations for the location of a cue within scenes (Rolls et al., 1989b), consistent with our view that the hippocampus itself may play a role in intermediate-term storage for relational representations (see above).
The absence of lasting representations for single stimuli in hippocampal neurons has led Brown and colleagues (1987) to question whether the hippocampus is involved in memory storage. Indeed, by this criterion, their point is well taken. However, there is now compelling neurophysiological evidence indicating that the hippocampus itself does indeed play a role in recognition memory, albeit one that does not include the storage of encodings for single specific items. Instead, it appears that hippocampal cellular activity at the time of recognition reflects judgements derived from comparisons between the sample and match cues, and that this matching process may contribute to recognition performance. For example, during stimulus sampling in DNMS tasks, hippocampal neurons in both rats (Figure 4c; Otto & Eichenbaum, 1992b; Sakurai, 1990) and monkeys (Riches et al., 1991) respond differentially to the "match" and "non- match" relationship between stimuli, and do so on all trials independent of the particular sensory stimuli that compose the stimulus comparisons. Thus the hippocampal representation does not reflect the perceptual qualities of particular stimuli but, rather, the abstraction of relevant relations among those stimuli. In other words, during recognition memory tasks, the responses of hippocampal cells depend on both previous and current stimuli, but their firing reflects only the outcome of the match or non-match judgement. We take such a neural correlate of abstract relationships as precisely the kind of representation that would be expected to support a relational memory organization. Relating these findings to descriptions of the relatively limited effects of selective hippocampal formation or fornix lesions on DNMS performance, it appears that this kind of processing may contribute to, but is not ultimately critical for, intermediate-term recognition memory. In addition, even in discrimination tasks where comparison of current stimuli to recently presented cues is not explicitly required, hippocampal neuronal activity reflects the processing of such comparisons (Eichenbaum et al., 1986; Foster et al., 1986). These findings confirm the similarity of hippocampal involvement in DNMS and discrimination learning described above.
Our characterization of hippocampal cellular activity as reflecting abstractions of the relations among cues is consistent with the observation that hippocampal cellular activity is associated with both an animal's position within an environment (O'Keefe, 1976) and spatial relationships among environmental cues and critical discriminative stimuli (Wiener et al., 1989; Wible et al., 1986; for review see Eichenbaum and Cohen, 1988). Spatial representations ("place fields") of hippocampal neurons do not depend on the immediate presence of any particular cue, and can sustain subtle alterations of the environment or removal of one of several spatial cues (O'Keefe and Conway, 1978; Hill and Best, 1981). Furthermore these spatial representations can be highly selective to particular environments (Thompson and Best, 1989), and spatial codings are unrelated across different environments and different behavioral paradigms performed in the same environment (Wiener et al., 1989; Cahusac et al., 1989). Environmental alterations that change the rat's judgement about the overall individuality of the environment result in dramatic and unpredictable effects on the hippocampal spatial representation (Breese et al., 1989; Muller and Kubie, 1987; Bostock et al., 1991). For example, after a stable spatial representation was observed in an environment, a subtle stimulus change at first produced no alteration in the spatial representation but, after multiple comparisons between the original and changed environment were permitted, a new and unpredictable spatial representation appeared for the changed environment (Bostock et al., 1991). Conversely, even when all the cues that ordinarily determine spatial orientation are removed, hippocampal place representations persist in correspondence with the animal's judgement of its location as reflected by behavioral choices (O'Keefe and Speakman, 1987; see also Muller et al., 1987; Quirk et al., 1990). Each of these findings is consistent with our conclusion that hippocampal processing does not involve sensory- specific codings, but rather reflects the outcome of relational judgements in terms of abstract, and in these cases spatial, relations in current or past experience. Finally, this integration of the findings on place cells with our conception of relational processing by the hippocampus is consistent with our notion that the critical role of the hippocampus in spatial memory is a strong example of its more general role in memory organizations (see above).
Finally, while there is little data that directly bears on this question, it is appropriate to ask how hippocampal representations of abstract relations can influence or mediate storage of stimulus specific representations in parahippocampal and cortical areas. Of particular relevance are recent findings of Miyashita et al. (1992) suggesting that neurons at the border of the parahippocampal region and temporal association cortex represent long term relational representations. The firing patterns of these cells reflected reliable repetitive sequences of visual stimuli presented in a recognition memory task and assigned relations among stimuli presented in a visual paired-associate task. Although it has not been demonstrated that these neural correlates were dependent on hippocampal function, the learning of sensory paired-associates has been shown to be sensitive to damage in the parahippocampal region (Bunsey and Eichenbaum, 1993). Thus, it may be that processing within the hippocampal system mediates a permanent cortical organization of perceptual and other cortical representations.
6. Combining the Anatomical, Behavioral, and Neurophysiological Findings in a Model of Memory Processing by the Cortex and Hippocampal System.
Accumulating evidence derived from several sources converge on a set characteristics for each stage of the model presented in Figure 1. Both the neuropsychological findings on amnesia and characterizations of the behavioral physiology of the hippocampal system serve to dissociate the temporal and representational properties or components of hippocampal-dependent memory processing, consistent with our view that these properties should be thought of as orthogonal dimensions. Moreover, the anatomical bases of this dissociation provide a preliminary assignment of differential processing functions to specific hippocampal areas contributing to declarative memory. Observations on the behavioral physiology of neocortical and hippocampal structures complement the neuropsychological findings and provide greater detail regarding the respective roles of specific components of this system.
The characteristics of these components are outlined in the following putative scenario for successive stages of memory processing. Prior to processing by the hippocampal system, neocortical areas create specific perceptual representations that can be sustained briefly. These representations are very sensitive to interference from intervening perceptual processing which likely results in new neural activity patterns that supplant such "active" memory representations. Neocortical areas are also capable of "passive" maintenance of memory traces in a form that may persist as long as the level of intervening interference is low. Such memory traces are seen to support perceptual matchings between current and stored representations and can support performance in short-term recognition and priming, consistent with the observed sparing of short term memory even in severe amnesia.
At the first stage of processing within the hippocampal system, perceptual codings reach the parahippocampal region where functionally distinct representations of the same events converge prior to processing in the hippocampal formation itself. These cortical areas do not hold information by sustaining neural activity, but can maintain a passive trace that persists through considerable interference and intervening processing. A matching between the intermediate-term store in the parahippocampal region and current representations in neocortex could mediate retention in recognition memory paradigms, accounting for the finding that parahippocampal areas may be sufficient to support DNMS performance in the absence of normal hippocampal function. However, in our view, the representation in the parahippocampal region does not constitute a full relational coding, and therefore would not be expected to be sufficient, even during intermediate periods, in tasks in which a relational representation or its flexible expression is required.
At the second stage of hippocampal system processing, the hippocampal formation does not maintain a memory representation of single sensory cues, but rather processes comparisons between current stimuli and representations of previous stim