Howard Eichenbaum
2004 Neuron  
cessing of associations among disparate elements of an experience (e.g., Squire and Zola-Morgan, 1991; Rolls, The hippocampus serves a critical role in declarative 1990; O'Reilly and Rudy, 2001; Ryan and Cohen, 2003; memory-our capacity to recall everyday facts and Morris et al., 2003) and spatial and temporal integration events. Recent studies using functional brain imaging of events into the record of experiences (e.g., Levy, in humans and neuropsychological analyses of hu-1989; Kesner, 1990;
more » ... Wallenstein et al., 1998; Redish, mans and animals with hippocampal damage have 1999; Burgess et al., 2002; Holscher, 2003) . Rather than revealed some of the elemental cognitive processes describing and comparing these alternative perspecmediated by the hippocampus. In addition, recent tives, this review will focus on the common themes characterizations of neuronal firing patterns in behavamong them, guided by a framework for conceptualizing ing animals and humans have suggested how neural cognitive processes that underlie declarative memory representations in the hippocampus underlie those eland prominent features of hippocampal circuitry and emental cognitive processes in the service of declaraplasticity. I will discuss three elemental cognitive protive memory. cesses that are mediated by the hippocampus-associative representation, sequential organization, and relational The domain of memory functions that are mediated by networking-and a set of characteristic properties of the hippocampus and the surrounding anatomically ashippocampal circuitry that could support these prosociated cortical regions was outlined nearly 50 years cesses: convergent afferents, recurrent connections, ago in the case study of H.M., a man who became amneand long-term potentiation (LTP). sic following removal of the medial temporal lobe to Cognitive Processes alleviate his epileptic seizures (Scoville and Milner, 1957; We live our lives through personal experiences, and our Corkin, 1984). H.M. is severely impaired in declarative initial construction of reality within consciousness is a memory, but his perceptual and cognitive abilities are form of episodic buffer that contains a representation intact, as are his capacities for working memory and of the stream of events as they just occurred (Baddeley, perceptual and motor skill learning. Furthermore, the 2000). Consider, for example, your episodic memory of deficit in new learning ability is accompanied by a tema day at a recent scientific meeting. You might recall a porally graded retrograde amnesia, such that H.M. can specific encounter with a colleague, your discussion recall information obtained remotely in life but he is about personal matters and about specific presentaimpaired in recalling events that occurred recently betions you and she/he had heard at the conference. Each fore the onset of amnesia. These observations suggest event within this episodic memory includes a combinathat the memory processing mediated by the hippocamtion of features: yourself, your colleague, what she/he pal system begins during learning and continues to conand you said, and where the conversation took place. tribute to the consolidation of memories over a pro-In addition, a vivid episodic representation for the enlonged period. counter is organized according to the order of events; Succeeding neuropsychological analyses on amnesic it unfolds as a "mental replay" of the encounter extended patients and functional imaging studies on normal huover time (Tulving, 2002) . Thus, that memory can be mans have elaborated the domain of capacities that deconstructed into a series of associative representaare dependent on the medial temporal region and, in tions, wherein each discrete event includes the relevant particular, on the hippocampus itself (Aggleton and people, their actions, and the place where that event Brown, 1999; Eichenbaum and Cohen, 2001; Squire et occurred, and these representations are sequentially oral., 2004). These studies have emphasized the critical ganized to compose the flow of events in that unique role of the hippocampus in two components of declaraexperience ( Figure 1 ). tive memory. The hippocampus plays a critical role in In addition, episodic memories do not exist in isolaepisodic memory, our capacity for the recollection of tion, but instead share many features with other memounique personal experiences; and the hippocampus is ries that bear common or closely related information. involved in particular aspects of the acquisition of se-You met that colleague on previous occasions and in mantic or factual knowledge. In addition to detailed different places, and you have discussed related sciencharacterizations of spared and impaired memory catific work on each of those occasions. During the current pacities in animals with selective hippocampal damage encounter, you can recall both specific previous discusand recordings of hippocampal neuronal activity in besions and general information that you have accrued having animals and humans, these studies have begun from many related experiences. A simple and effective to reveal how the hippocampus mediates these complex way to organize both the specific (episodic) memories memory functions. and the common (semantic) information accrued across memories is to encode common features of related experiences into the same representational elements (Fig-*Correspondence: hbe@bu.edu Neuron 110 CA3 population, sparse, and involve mainly excitatory glutamatergic synapses (Amaral and Witter, 1995; Treves and Rolls, 1994). (3) The hippocampus is noted for the prevalence of rapid synaptic plasticity, known as longterm potentiation (Bliss and Collingridge, 1993). In particular, a form of LTP that is dependent on N-methyl D-aspartate (NMDA) receptors has been strongly linked to memory (Martin et al., 2000) and to the memoryassociated firing properties of hippocampal neurons (Shapiro and Eichenbaum, 1999). These properties can support the elemental cognitive Figure 1. Schematic Diagram of a Simple Relational Network processes that underlie declarative memory. First, asso-The network is composed of two episodic memories (A and B), ciative representations can be formed by simultaneous each construed as a sequence of elements (1-6) that represent the conjuction of an event and the place where it occurred. C is an activation of multiple afferents to CA3 principal cells, element that contains the same features in both episodes. D is an leading to LTP of weakly as well as strongly activating element that contains only some of the common information. inputs ( Figure 2A ). This associative LTP supports pattern completion, such that subsequent presentation of any part of the representation would fire the cell, constituting ure 1). These elements organize memories into a relaretrieval of the entire association (for review, see Nakational network that offers two potentially quite useful zawa et al., 2004). In addition, recurrent connections properties. First, memories for the common features can support pattern completion by spreading the activabecome "timeless" semantic elements, not bound to any tion of some elements to all elements that compose particular episode in which they were acquired. Second, a previously activated CA3 network (McNaughton and the elements that encode common features link memo-Morris, 1987; Treves and Rolls, 1994). ries to one another, allowing one to compare and con-Second, several recent computational models have trast memories and to make inferences among indirectly emphasized the sequential organization of memory reprelated events. These properties of relational memory resentations supported by the hippocampus and, in parrepresentation underlie the hallmark "flexibility" of deticular, area CA3. Levy (1989 Levy ( , 1996 proposed that clarative memory expression (Cohen, 1984) . Also, the unique characteristics of hippocampal area CA3, specifcontinued abstraction of common features and linking ically the sparseness and largely excitatory nature of its of memories may underlie a more extensive association recurrent connectivity, combined with rapid synaptic and interleaving of memories during the prolonged peplasticity, inherently produces asymmetric connections riod of consolidation (Ribot, 1882; Burnham, 1903; that can represent sequences of information from a sin-McClelland et al., 1995). gle patterned input and can spontaneously reproduce Hippocampal Circuitry and Plasticity learned sequences. Thus, according to these models, The elemental cognitive processes of declarative memwhen temporally patterned inputs reach the hippocamory introduced above could be mediated by a combinapus, a rapid LTP mechanism enhances connections betion of three well-known properties of hippocampal cirtween cells that fire in sequence. The likelihood of a cuitry; these properties are particularly prominent in reciprocal (backward) connection of forward connected region CA3 (Figure 2 ). (1) The hippocampus receives cells is very low due to the overall sparse connectivity. convergent afferents from virtually all cortical associa-Therefore, the enhancement of recurrent connections is tion areas, and these inputs are widely distributed onto mostly unidirectional, leading to an asymmetry of the the cell population in multiple subdivisions of the hippoenhanced connectivity. When partial inputs are reprocampus (Amaral and Witter, 1995). Thus, CA3 principal duced, the network is more likely to complete the secells have as their main afferents considerable highquence of the full initial input pattern. In addition, when level perceptual information about attended stimuli and sequences are repeated (practiced) just a few times, spatial cues as well as other information from diverse cells that represent constant background information cortical regions. (2) The principal neurons of area CA3 provide a context that bridges the firings of neurons send considerable projections to other CA3 principal representing salient sequential events (Levy, 1989, 1996; Wallenstein and Hasselmo, 1997; Wallenstein et al., cells. These recurrent connections are broad across the Figure 2. Schematic Diagram of the Circuitry of the Principal Cells in Hippocampal Area CA3 that Could Mediate Properties of Relational Networks Two major pathways are shown, one pathway involving inputs from diverse neocortical association areas that send widespread projections onto the dendrites of pyramidal cells and the other pathway involving recurrent projections from pyramidal cells projecting sparsely onto other pyramidal cells in the same network. Connections of both of these pathways are subject to activity-dependent rapid synaptic plasticity. Review 111 1998). Although sequence storage and recall were ini-was activated during correct recollections that included perceptual details or associations made with a word tially proposed as characteristics of CA3, subsequent models (Lisman, 1999) have shown how more complex during a previous study phase as compared to a sense of familiarity with the words. In a study that directly and reciprocally connected recurrent networks in the dentate gyrus and CA3 can coordinate to provide more compared the encoding of stimuli in combination or separately, Henke et al. (1997) observed greater hippo-faithful recall of sequences, leaving CA1 to decode the sequence signals back to the cortex and to compare campal activation when subjects associated a person with a house, as compared to making independent judg-predictions of the network for the next sequential item to actual information as it arrives. ments about the person and house. Also, Giovanello et al. (2003a) reported greater hippocampal activation Third, the same computational models that emphasize temporal organization in episodic memory representa-when subjects recognized previously presented word pairings as compared to rearranged word pairs or words tions provide a mechanism for linking distinct memory representations into a relational network. These net-studied separately. Davachi and Wagner (2002) found that the hippocampus is activated during the encoding works include cells that receive no external inputs but develop firing patterns that are regularly associated with of multiple items and is more activated when subjects are required to link the items to one another by system-a particular sequence or with overlapping sequences (Levy, 1996; Sohal and Hasselmo, 1998; Wallenstein et atic comparisons as compared to rote rehearsal of individual items. Furthermore, the magnitude of hippocam-al., 1998). In the situation where episodes are repeated, these cells provide a local temporal context in which pal activation during the comparison and linkage of the items predicted later success in recognition. Other re-items within a particular sequence are linked together. When these links incorporate events that are unique cent studies have revealed activation within subfields of the hippocampus during the encoding of face-name to a particular episode, they can assist the network in disambiguating successive patterns in overlapping but associations (Zeineh et al., 2003) and along the entire longitudinal extent of the hippocampus when subjects distinct sequences. At the same time, when the links are activated similarly by separate episodes that share studied face-name pairs that they later remembered with high confidence (Sperling et al., 2003). Also, Small a series of overlapping features, they can allow the association of discontiguous episodes that share those fea-et al. (2001) reported that, whereas viewing faces and names independently activated separate areas within tures. Thus, the same network properties that support the encoding of episodes as sequences of events also the hippocampus, the viewing of faces and names in combination activated a distinct area bridging between contain the means to link and disambiguate related episodes (see Agster et al., 2002, for an experimental test). the separate face and name activations; and the latter area was also activated during recall of names when Convergent inputs, recurrency, and LTP are, to differing extents, also properties of other subdivisions of cued by the face. Other studies report activation of the hippocampus the hippocampus as well as cortical areas (e.g., Bear, 1996). Therefore, the functional mechanisms described during the retrieval of autobiographical experiences (see review in Maguire, 2001). Maguire et al. (2000) reported here may contribute to memory processing throughout the hippocampus and may support the permanent con-selective activation in the medial temporal region during the retrieval of multiple aspects of autobiographical solidation of memories within widespread areas of the cerebral cortex (e.g., Frankland et al., 2004; Maviel et events but not the retrieval of public events. Addis et al. (2004) also found that the level of detail, personal al., 2004). The special role of these mechanisms in the hippocampus may be derived from the high extent of significance, and emotionality each contributed to hippocampal activation of autobiographical memories. The diverse convergent inputs, the strong recurrency, and the rapidity of LTP that are prominent within hippocam-involvement of the hippocampus in processing complex material is not limited to autobiographical details but pal circuitry. Are the elemental cognitive processes of declarative extends broadly, for example, to recollection of the context of learning in formal tests of memory (e.g., Davachi memory that are highlighted in this framework actually mediated by the hippocampus? Do hippocampal neural et al., 2003; Ranganath et al., 2003). Some studies have provided evidence that different representations reflect these cognitive processes? The following sections consider the evidence from neurocomponents of the medial temporal lobe may play dispsychological and physiological studies regarding each tinct roles in memory. Gabrieli et al. (1997) showed subof the three elemental cognitive processes of declarative jects pictures that contained line drawings of common memory introduced above. objects and animals. When subjects were again shown the pictures and recalled their names, a part of the hippocampus was selectively activated. In contrast, as had Associative Representations previously been reported by Stern et al. (1996), when There is a growing body of evidence that suggests that subjects were shown novel or highly familiar pictures, the hippocampus encodes associations among stimuli, the parahippocampal gyrus was activated. Similarly, Daactions, and places that compose discrete events. Sevvachi et al. (2003) presented subjects with words and eral functional imaging studies have examined whether asked them either to imagine a spatial scene described the hippocampus is more activated during the encoding by the word (e.g., a garbage dump for "dirty") or to read or retrieval of associations among many elements of a the word backward. They found that the hippocampus memory-a characteristic of context-rich episodic was selectively activated when subjects successfully memories ( for review, see Cohen et al., 1999). For example, Eldridge et al. (2000) found that the hippocampus encoded the contextual cues, whereas the surrounding Neuron 112 cortex was selectively activated during successful en-manipulated independently of the spatial position of an object. In one study, rats were initially exposed to two coding of the superficial qualities of single words. On the other hand, in several of the other studies objects in particular places in one of two environmental chambers (Mumby et al., 2002). In subsequent recogni-described above, cortical areas adjacent to the hippocampus, as well as additional cortical areas, were also tion testing, the place of the object or the context was changed. Normal rats show increased exploration of activated during associative processing. In particular, one recent study emphasized that the hippocampus and objects moved to new places or in novel contexts. By contrast, rats with hippocampal damage failed to recog-adjacent cortex showed equivalent increased activation for associative over nonassociative information (Kirwan nize objects when either the place or contextual information was changed. Another study failed to find a deficit and Stark, 2004). These findings are consistent with the view that several areas of the cerebral cortex also in object-place recognition following hippocampal disconnection, but the same animals had a severe deficit process combinations of stimuli, albeit in a different type of processing than that mediated by the hippocampus in recognition based on a combination of contextual and place cues (Eacott and Norman, 2004). Perhaps the (Eichenbaum, 1994; O'Reilly and Rudy, 2001; Bussey et al., 2002). Thus, while associative processing may be strongest evidence that the hippocampus is critical for learning the context of important events comes from prominent within the hippocampus, its specific role in representing associations is not clarified by functional studies of fear conditioning (Phillips and LeDoux, 1992). These studies are based on the conditioning protocol imaging (Squire et al., 2004). Consistent with the notion that the hippocampus is in which a tone and shock are paired repeatedly such that rats become fearful of the tone. In addition to a critical for forming associations among items in events, recent neuropsychological studies have found that rec-conditioned fear for the tone cue, rats also become fearful of the context in which the tones and shock were ognition of associations is impaired even when recognition for single items is spared in amnesic patients ( Giova-presented, evidenced as freezing behavior and other indices of fear when the animal is returned to the envi-nello et al., 2003b; Turriziani et al., 2004). These studies reported impairment in recognition memory for associa-ronment where conditioning occurred. Damage to the hippocampus eliminates the contextual fear condition-tions between words or between faces or face-occupation pairs as compared to normal performance in the ing without affecting conditioned fear to the tone. Finally, the evidence concerning the specific involve-recognition of single items. On the other hand, Stark et al. (2002) reported that hippocampal damage resulted ment of the hippocampus in recognition for single items assessed within the delayed nonmatch-to-sample in impairment in both associative and single-item recognition, highlighting the need to clarify the nature of asso-(DNMS) test is not clear. In monkeys, some studies report a consistent partial deficit in DNMS performance ciative information that composes an "event." The mixture of findings from both functional imaging in normal (Zola et al., 2000), whereas other studies report no deficit (Murray and Mishkin, 1998). The source of this discrep-subjects and from amnesic patients suggests that studies on humans have not conclusively distinguished the ancy is under debate (Baxter and Murray, 2001; Zola and Squire, 2001). By contrast, damage to the cortical specific role of the hippocampus in stimulus association from that of adjacent cortical areas in the same system areas surrounding the hippocampus produces a consistent severe impairment in DNMS (Zola-Morgan et al., (Squire et al., 2004). In animals, there is substantial evidence that selective 1989; Meunier et al., 1993). Also, the same animals with selective hippocampal damage that show relatively hippocampal damage results in deficits in forming a memory for the context or location where items were spared DNMS performance are severely impaired on recognition assessed by novel item exploration (Zola previously experienced. Gaffan (1994) trained monkeys on a set of discrimination problems composed as ob-et al., 2000; Nemanic et al., 2004). In rats, reports of impairment or spared performance on DNMS following jects stimuli presented on a computer screen with different kinds of background patterns. Animals with the hip-selective hippocampal damage are also variable, whereas damage to the surrounding cortical areas consistently pocampus disconnected learned object discriminations at the normal rate when the background simply varied results in a severe deficit (reviewed in Steckler et al., 1998; Mumby, 2001). and did not predict the location of the objects or their reward values. By contrast, the same animals were most One possible explanation for the different findings on selective hippocampal damage is that there may be two impaired when the background context predicted the location of the rewarded object. Day et al. (2003) initially types of memory that contribute to recognition performance, only one of which is affected by hippocampal allowed rats to find different flavored rewards at specific locations in an open platform, then tested their memory damage (Brown and Aggleton, 2001). This hypothesis was recently addressed using signal detection tech-for the location of those events by providing an additional flavored reward associated with one of the loca-niques to distinguish contributions of recollection of associative information in prior experiences versus a sense tions. After a single exposure to the flavor in a particular location, animals could return to the location where a of familiarity for the previously experienced items, as has been done in humans (Yonelinas, 2001, 2002). Rats flavor had previously been consumed when cued by the flavor alone. By contrast, inactivation of the hippocam-were trained on a variant of DNMS in which they initially sampled a series of odors and then judged old and new pus or blockade of NMDA receptors prevented encoding of the flavor-place association. test stimuli across a range of response criteria (Fortin et al., 2004). Rats with hippocampal damage showed Evidence indicating hippocampal involvement in processing contextual information in recognition memory the typical partial deficit in overall recognition performance. An analysis of receiver operating characteristics comes from studies where the overall environment was Review 113 (ROC) indicated that, in normal rats, both recollection gradient of recognition performance in rats with hippocampal damage indicated that they had normal access and familiarity contributed to recognition. By contrast, in to the differences in trace strengths for the odors. Yet rats with selective hippocampal damage, ROC analysis these intact trace strength differences were not suffirevealed that recognition was entirely supported by facient to support consistent discrimination on the order miliarity. These results can explain the overall partial probe tests. These considerations strongly suggest that deficit in DNMS observed across species and suggest normal rats also could not utilize the relative strengths that the mixed findings across recognition tasks may of memories for the recently experienced odors and be a consequence of the differential demands for recolinstead based their order judgments directly on rememlection and familiarity across tasks in animals and in bering the odor sequence. Contrary to the argument humans (Pascalis et al., 2004) . that animals lack episodic memory because they are "stuck in time" (Roberts, 2002; Tulving, 2002), these Sequential Organization observations suggest that animals have the capacity to A property of episodic memory that is prominent in comrecollect the flow of events in unique experiences. putational modeling involves the organization of an epi-Another major challenge for a robust model of episode as a sequence of events that unfolds over time. sodic memory is the requirement for a capacity to develop These studies focus on the observation that vivid epirepresentations that can distinguish two experiences that sodic memories contain not only a particular item or share common elements. Levy (1996) proposed that memitems that one is attempting to recall, but also the experiory for the ordering of events mediated by the hippoence of events that precede and follow (Tulving, 2002). campus may be especially important when the event A consideration of memory for the orderliness of events sequences have overlapping elements through which in unique experiences, a capacity that can be tested in memory of earlier elements must be remembered to animals, may provide a fruitful avenue for neurobiologicomplete each distinct sequence. In order to test cal explorations of episodic memory. For example, whether sequence disambiguation is a fundamental fea- Honey et al. (1998) provided a simple demonstration ture of memory processing dependent on the hippocamof the importance of temporal order in hippocampal pus, Agster et al. (2002) trained rats on a sequence processing, reporting that hippocampal lesions disdisambiguation task designed after Levy's (1996) formal rupted animals' normal orienting response when a pair model that involved two series of events that overlap in of stimuli were presented in the opposite order of previthe middle items. The sequences were presented as ous exposures. a series of six pairwise odor choices where, for each To investigate the specific role of the hippocampus sequence, selection of the appropriate odor at each in remembering the order of a series of events in unique choice point was rewarded. Each trial began with two experiences, recent studies have employed a behavioral forced choices that initiated the production of one of protocol that assesses memory for episodes that are the two sequences. Then the animal was presented with composed of a unique sequence of olfactory stimuli two forced choices that were the same for both se-(Fortin et al., 2002; see also Kesner et al., 2002). In one quences. Subsequently, the subject was allowed a free of these studies, memory for the sequential order of choice and was rewarded for selecting the odor asodor events was directly compared with recognition of signed to the ongoing sequence. Normal rats maintained the odors in the list independent of memory for their a high level of performance on the critical free choice, order. On each trial, rats were presented with a series indicating that they could remember the beginning of of five odors that were selected randomly from a large the sequence. By contrast, the rats with hippocampal pool of common household scents. Memory for each damage performed poorly on the free choice either when series was subsequently probed using a choice test sequences were alternated rapidly (high proactive interwhere the animal was reinforced for selecting the earlier ference) or when a delay was imposed before the critiof two of the odors that had appeared in the series or cal choice. (in later testing) was reinforced for selecting a novel These results are further clarified by a study that exodor against one that had appeared in the series. Normal amined spatial sequence learning in rats with damage rats performed both tasks well. Rats with hippocampal to the hippocampus or striatum. DeCoteau and Kesner lesions could recognize items that had appeared in the (2000) developed two versions of a task where rats enseries but were severely impaired in judging their order. tered a series of six arms of a radial maze in which A potential confound in any study that employs time as successive choices were controlled by a door to each a critical dimension in episodic memory is that memories arm. Striatal lesions resulted in impairment on an "imobtained at different times are likely to differ in the plicit" version of the task where rats simply ran through strength of their memory traces due to the inherent decthe fixed sequence of arms opened in order. By contrast, remental nature of memory traces. To what extent could hippocampal lesions resulted in impairment on an "exnormal animals be using differences in the relative plicit" version of the task where rats had to stand in strengths of memory traces for the odors to judge their front of the door to each succeeding arm before it sequential order? The observation of a temporal gradiopened. These findings are consistent with the observaent in recognition performance by normal animals sugtion that sequence learning can be mediated by declaragests that memories were in fact stronger for the more tive and nondeclarative strategies involving distinct recently presented items in each sequence. These difmemory systems (Keele et al., 2003), and these forms ferences in trace strength potentially provide sufficient of representation are independent (Reber and Squire, signals for the animals to judge the order of their presen-1998). Furthermore, in humans the hippocampus is implicated in mediating a declarative representation of se-tation. However, the observation of the same temporal Neuron 114 quence memory that can be established independently nated 87% of famous from fictitious names and provided factual details for a third of the recognized people. Simiof conscious recollection (Schendan et al., 2003; Keele larly, H.M. also shows some spared acquisition of spatial et al., 2003), providing a strong parallel to the observaknowledge in that he can draw a reasonably accurate tions on rats. floor plan of the home in which he lived from 5 years to 21 years after he became amnesic (Corkin, 2002). Linking Memories within Relational Networks Does the partial deficit in semantic memory following The third major feature of the proposed model involves hippocampal damage reflect a quantitative difference the linking of episodic memories into relational networks in an overall declarative memory capacity of the medial in order to abstract the common features among related temporal lobe system (Manns et al., 2003) or full and memories and to mediate flexible memory expression. selective loss of some quality of semantic processing? Within this model, the hippocampus contributes to se-McClelland et al. (1995) and, more recently, O'Reilly and mantic memory by the construction of relational net-Rudy (2001) proposed that the hippocampus and cortiworks that coordinate memories stored in the cerebral cal areas both have powerful learning capacities for cortex (Eichenbaum et al., 1999). From this perspective, learning stimulus combinations that compose declarasemantic memories are not directly mediated by the tive knowledge. However, they argue, these systems hippocampus; rather, the hippocampus links memories differ in their information processing mechanisms, such in support of the flexibility of their expression through that the hippocampus rapidly learns about individual comparisons and generalizations across memories. experiences and prevents interference by separating Therefore, one might expect conditions in which semanrepresentations of those experiences, whereas the cortic memory might be possible without requisite hippotex gradually extracts regularities over many expericampal involvement, albeit such memories might differ ences. This view suggests no ultimate difference in the in the available range of their expression. nature of memory representations with or without hippo-Indeed, some recent studies have emphasized the campal participation but, rather, differences in how relative sparing of semantic memory in amnesia associmemories are formed and in the success rate following ated with selective damage to the hippocampal region only a single experience. (e.g., Vargha-Khadem et al., 1997; Verfaellie et al., 2000; To address this issue, Bayley and Squire (2002) trained O'Kane et al., 2004). However, there is also substantial a patient (EP) with extensive medial temporal lobe damevidence that the normal acquisition and flexible expresage on factual information by presenting novel threesion of acquired semantic memories are very much deword sentences repeatedly in several sessions across pendent on the hippocampal region. Manns et al. (2003) three months. EP learned more slowly than controls but examined the severity of retrograde and anterograde showed clear gradual improvement in both cued recall amnesia for semantic information in patients with damand recognition of the last word in learned sentences. age limited to the hippocampus and adjacent entorhinal Furthermore, unlike control subjects, EP never conarea. These patients were impaired in memory for news sciously recalled the items, showing no increased confievents for a year or more prior to the onset of amnesia dence for correct choices and no difference in response and were impaired for information obtained after they times for correct and incorrect responses, both of which became amnesic. They performed well in recognizing typically accompany conscious recollection in normal names of famous people but poorly in judging whether subjects. Also, EP failed when a target word was rethese people were living or deceased, and this deficit placed with a synonym, showing that, unlike normal was apparent even when corrected for names that subjects, EP's memory did not emphasize the semantic evoked episodic memory in normal subjects. Complecontent of the sentences. These findings are consistent mentary evidence from functional brain imaging studies with previous reports of spared capacities of amnesic has revealed that a large cortical network, including patients in learning computer language routines (Glisky the hippocampus, is activated during the acquisition of et al., 1986) in which successful performance is rigidly everyday factual information as compared to a baseline tied to the circumstances of initial learning. These obsertask that included similar episodic experience but not vations indicate that spared semantic memory in amnegeneral knowledge (Maguire and Frith, 2004). sia is qualitatively different than that of normal subjects, Still, the patients in Manns and colleagues' study leading Bayley and Squire (2002) to characterize the showed evidence of the acquisition of substantial sepreserved verbal learning abilities as nondeclarative. mantic knowledge, albeit less than that of normal sub-Animal models have also provided clues about how jects. Even in cases of severe amnesia following exteninformation processing by the hippocampus might exsive damage to the hippocampus and surrounding plain the findings on semantic memory in amnesic pacortical areas, there is compelling evidence of some tients. In particular, some studies have focused directly degree of spared semantic learning. Although several on the learning of multiple related problems and their studies have shown the relative sparing of semantic integration into networks of memory that support fleximemory in amnesia, the most impressive of these is ble, inferential judgments. One study compared the abilthe O'Kane et al. (2004) report that patient H.M., while ity of normal rats and rats with selective damage to impaired compared with normal control subjects, shows the hippocampus on their ability to learn a set of odor a remarkable amount of acquired knowledge about peoproblems and to interleave the representations of these ple who become famous following the onset of his amneproblems in support of novel inferential judgments sia. When prompted with a first name, H.M. recalled the (Bunsey and Eichenbaum, 1996). Animals were initially last name of 12 out of 35 famous personalities. Furthertrained on two sets of overlapping odor paired associates (e.g., A goes with B, B goes with C). Then the rats more, in a forced choice recognition task, H.M. discrimi-Review 115 were given probe tests to determine if they could infer described results provide compelling evidence of hippothe relationships between items that were only indirectly campal involvement across species in the flexible exassociated through the common elements (A goes with pression of memories using relational networks. C?). Normal rats learned the paired associates and Combined, these findings suggest that, whereas the showed strong transitivity in the probe tests. Rats with cerebral cortex can mediate that acquisition of complex selective hippocampal lesions also learned the pairs stimulus conjunctions that compose semantic knowlover several trials but were severely impaired in the edge, the hippocampus performs an additional type of probes, showing no evidence of transitivity. processing that contributes to the networking of cortical In another experiment, rats learned a hierarchical sememories. The results on the transitive inference pararies of overlapping odor choice judgments (e.g., A Ͼ B, digm indicate that the hippocampus plays a critical role B Ͼ C, C Ͼ D, D Ͼ E) and then were probed on the in linking related memories according to their common relationship between indirectly related items (B Ͼ D?). features, and this linkage results in a network that can Normal rats learned the series and showed robust transisupport inferences between items in memory that are tive inference on the probe tests. Rats with hippocampal only indirectly related. Such performance suggests how damage also learned each of the initial premises but a general networking of memories can underlie the flexifailed to show transitivity (Dusek and Eichenbaum, bility of declarative memory expression. 1997). The combined findings from these studies show that rats with hippocampal damage can learn even com-Observations on the Firing Patterns plex associations, such as those embodied in the odor of Hippocampal Neurons paired associates and conditional discriminations. But, Additional evidence on the nature of information prowithout a hippocampus, they do not interleave the discessing by the hippocampus is provided by charactertinct experiences according to their overlapping eleizations of the firing patterns of hippocampal neurons ments to form a relational network that supports inferenin animals and humans performing memory tasks. Obtial and flexible expression of their memories. servations from rats, monkeys, and humans and across Recently, the same findings were reported in a study many different behavioral protocols show that hippoon monkeys (Buckmaster et al., 2004) . The same animals campal neuronal activity reflects each of the three funwere trained on both transitivity problems described damental features of declarative memory discussed above, using visual instead of olfactory stimuli. Monkeys above: the combinations of event features, the sequencwith lesions of the entorhinal cortex learned visual paired ing of events in specific experiences, and the represenassociates, albeit more slowly than normal monkeys. tation of common features of events that could link dis-However, whereas normal monkeys showed strong trantinct memories. sitivity on indirect associations, monkeys with entorhinal Associative Representations lesions did not. Also, normal monkeys and monkeys A large body of evidence shows that hippocampal neuwith entorhinal lesions learned a series of overlapping rons encode an animal's location within its environment discrimination problems as rapidly as the controls. Nor- (Muller et al., 1999; Best et al., 2001). In addition, howmal monkeys also showed robust transitive inference ever, many studies have shown that hippocampal neuon probe trials, but the monkeys with entorhinal lesions rons fire associated with the ongoing behavior and the performed no better than chance. context of events as well as the animal's location (Eich-Complementary evidence on the role of the hippocamenbaum et al., 1999). The combination of spatial and pus in the networking of memories comes from two nonspatial features of events captured by hippocampal recent studies indicating that the hippocampus is selecneuronal activity is consistent with the view that the tively activated when humans make inferential memory hippocampus encodes many features of events and the judgments. In one study, subjects initially learned to places where they occur. associate each of two faces with a house and, sepa-Two recent studies highlight the associative coding rately, learned to associate pairs of faces (Preston et of events and places by hippocampal neurons. In one al., 2004). Then, during brain scanning, the subjects were study, rats were trained on an auditory fear conditioning tested on their ability to judge whether two faces that task (Moita et al., 2003) . Prior to fear conditioning, few were each associated with the same house were therehippocampal cells were activated by an auditory stimfore indirectly associated with each other and on ulus. Following pairings of tone presentations and whether they could remember trained face pairs. The shocks, many cells fired briskly to the tone when the hippocampus was selectively activated during the peranimal was in a particular place where the cell fired formance of the inferential judgment about indirectly above baseline. Another recent study examined the firrelated faces as compared to during memory for trained ing properties of hippocampal neurons in monkeys perface-house or face-face pairings. In the other study, forming a task where they rapidly learned new scenesubjects learned a series of choice judgments between location associations (Wirth et al., 2003) . Just as the pairs of visual patterns that contained overlapping elemonkeys acquired a new response to a location in the ments, just as in the studies on rats and monkeys, and scene, neurons in the hippocampus changed their firing as a control they also learned a set of nonoverlapping patterns to become selective to particular scenes. These choice judgments (Heckers et al., 2004) . The hippocamscene-location associations persist even long after pus was selectively activated during transitive judglearning is completed ( Yanike et al., 2004). ments as compared to novel nontransitive judgments. Wood et al. (1999) directly compared spatial and non-Under some circumstances, it may be possible to indispatial coding by hippocampal neurons by training anirectly relate items without a relational network (O'Reilly and Rudy, 2001; Van Elzakker et al., 2003), but the above mals to perform the same memory judgments at many Neuron 116 locations in the environment. Rats performed a task in episode to guide the choice on the current trial, and in that way, the task is similar in demands to those of which they had to recognize any of nine olfactory cues was placed in any of nine locations. Because the loca-episodic memory. If hippocampal neurons encode each sequential behavioral event and its locus within one type tion of the discriminative stimuli was varied systematically, cellular activity related to the stimuli and behavior of episode, then most cells should fire only when the rat is performing within either the left turn or the right could be dissociated from that related to the animal's location. A large subset of hippocampal neurons fired turn type of episode. This should be particularly evident when the rat is on the "stem" of the maze, when the rat only associated with a particular combination of the odor, the place where it was sampled, and the match/ traverses the same locations on both types of trials. Indeed, virtually all cells that fired when the rat was on nonmatch status of the odor. Similarly, Ekstrom et al. (2003) recorded the activity of hippocampal neurons the maze stem fired differentially on left turn versus right turn trials. The majority of cells showed strong in human subjects as they played a taxi driver game, searching for passengers picked up and dropped off at selectivity, some firing at over ten times the rate on one trial type, suggesting they were part of the representa-various locations in a virtual-reality town. Many of these cells fired selectively associated with specific combina-tions of only one type of episode. Other cells fired substantially on both trial types, potentially providing a link tions of a place and the view of a particular scene or a particular goal. Hippocampal cells that represent spe-between left turn and right turn representations by the common places traversed on both trial types. cific salient objects in the context of a particular environment have also been observed in studies of rats en-Recently, Ferbinteanu and Shapiro (2003) also reported that many hippocampal neurons fire associated gaged in foraging (Gothard et al., 1996; Rivard et al., 2004) and escape behavior (Hollup et al., 2001) in open with serial locations occupied as rats traverse different routes within the same environment. Furthermore, they fields. Thus, in rats, monkey, and humans, a prevalent property of hippocampal firing patterns involves the rep-modified the task to distinguish whether the activity of these cells reflected the immediate past experience of resentation of unique associations of stimuli, their significance, specific behaviors, and the places where these the animal (retrospective coding) or predicted its future path (prospective coding). They found that these cells events occur. Sequences of Events encode both past events and future goals of each route, with some cells encoding both kinds of information on Another common observation across species and many different behavioral protocols is that different hippo- the same trials. Furthermore, retrospective and prospective coding diminished on error trials, and some campal neurons become activated during virtually every moment of task performance, including during simple cells fired associated with the intention to proceed to a particular location even when a detour was required. behaviors such as foraging for food (e.g., Muller et al., 1987) as well as learning-related behaviors directed at These findings indicate that the overall hippocampal representations are neither retrospective or prospective relevant stimuli that have to be remembered (e.g., Hampson et al., 1993) . This general pattern is also observed per se and do not necessarily capture the precise details of behaviors or places that distinguish qualitatively simi-in a broad range of learning protocols, from studies that involve classical conditioning, discrimination learning, lar episodes. Rather, the hippocampal network encodes routes through space as a meaningful sequence of and nonmatching or matching to sample tasks to tests and a variety of maze tasks (for review, see Eichenbaum events that characterize a particular spatially extended experience. Additionally, the processing of previous et al., 1999). In each of these paradigms, animals are repeatedly presented with specific stimuli and rein-spatial experiences as the sequential activation of places may continue offline for a substantial period forcers, and they execute appropriate cognitive judgments and conditioned behaviors. Corresponding to (Louie and Wilson, 2001; Lee and Wilson, 2002; Nadasdy et al., 1999). each of these regular events, many hippocampal cells show time-locked activations associated with each se- Relational Networks In virtually all of the studies described above, some quential event. Also, as described above, many of these cells show striking specificities corresponding to partic-hippocampal neurons encode features that are common among different experiences-these representations ular combinations of stimuli, behaviors, and the spatial location of the event. Thus, within the overall network, could provide links between distinct memories. In Moita et al. (2003) study of auditory fear conditioning, whereas cellular activity can be characterized as a sequence of firings representing the step-by-step events in each some cells only fired to a tone when the animal was in a particular place, others fired associated with the tone behavioral episode. Furthermore, these sequential codings can be envi-wherever it was presented across trials. In the Wood et al. (1999) study on odor recognition memory, whereas sioned to represent a series of events and their places that compose a meaningful episode, and the information some cells showed striking associative coding of odors, their match/nonmatch status, and places, other cells contained in these representations both distinguishes and links related episodes. Recent studies on the spatial fired associated with one of those features across different trials. Some cells fired during a particular phase firing patterns of hippocampal neurons provide compelling data consistent with this characterization. In one of the approach toward any stimulus cup. Others fired differentially as the rat sampled a particular odor, re-study, rats were trained on the classic spatial alternation task in a modified T maze ( Wood et al., 2000; see also gardless of its location or match/nonmatch status. Other cells fired only when the rat sampled the odor at a partic-Frank et al., 2000). Performance on this task requires that the animal distinguish left turn and right turn epi-ular place, regardless of the odor or its status. Yet other cells fired differentially associated with the match and sodes and that it remember the immediately preceding Review 117 nonmatch status of the odor, regardless of the odor or the framework outlined at the outset of this review. The hippocampus is envisioned as critically involved in the where it was sampled. Similarly, in the Ekstrom et al. (2003) study on humans performing a virtual navigation rapid encoding of events as associations among stimulus elements and context, in the encoding of episodes task, whereas some hippocampal neurons fired associated with combinations of views, goals, and places, as sequences of events, and in linking episodes by common features into relational networks that support other cells fired when subjects viewed particular scenes, occupied particular locations, or had particular goals in flexible inferential memory expression. The details of memory representations are likely contained within finding passengers or locations for drop off. Also, in the Rivard et al. (2004) study of rats exploring objects in widespread areas of the cortex, such that the properties of declarative memory described here involve a combi-open fields, whereas some cells fired selectively associated with an object in one environment, others fired nation of hippocampal and cortical processing. Furthermore, repeated activation of the hippocampal-cortical associated with the same object across environments. The notion that these cells might reflect the linking of network during rehearsal, recall, and new related experiences, as well as during offline periods, could provide important features across experiences and the abstraction of common (semantic) information was highlighted the basis for a prolonged period of organization and consolidation of memories within the cerebral cortex in recent studies on monkeys and humans. Hampson et al. (2004) trained monkeys on matching-to-sample (Buzsaki, 1996; Eichenbaum, 2000; Lee and Wilson, 2002). problems, then probed the nature of the representation of stimuli by recording from hippocampal cells when the
doi:10.1016/j.neuron.2004.08.028 pmid:15450164 fatcat:anhtlasphjfbvej6a4ert6um7e