Aggleton, John P. & Brown, Malcolm W. (1999) Episodic memory, amnesia, and the hippocampal-anterior thalamic axis. BEHAVIORAL AND BRAIN SCIENCES (1999) 22(3)

Episodic memory, amnesia, and the hippocampal-anterior thalamic axis

Abstract:

Using new information from both clinical and experimental (lesion, electrophysiological, and gene-activation) studies with animals, the anatomy underlying anterograde amnesia is reformulated. The distinction between temporal lobe and diencephalic amnesia is of limited value because a common feature of anterograde amnesia is damage to part of an "extended hippocampal system" comprising the hippocampus, the fornix, the mamillary bodies and the anterior thalamic nuclei. This view, which can be traced back to Delay and Brion (1969), differs from other recent models in placing critical importance on the efferents from the hippocampus via the fornix to the diencephalon. These are necessary for the encoding and, hence, the effective subsequent recall of episodic memory. An additional feature of this hippocampal - anterior thalamic axis is the presence of projections back from the diencephalon to the temporal cortex and hippocampus that also support episodic memory. In contrast, this hippocampal system is not required for tests of item recognition that primarily tax familiarity judgements. Familiarity judgements reflect an independent process that depends on a distinct system involving the perirhinal cortex of the temporal lobe and the medial dorsal nucleus of the thalamus. In the large majority of amnesic cases both the hippocampal - anterior thalamic and the perirhinal - medial dorsal thalamic systems are compromised, leading to severe deficits in both recall and recognition.

Keywords:

amnesia, memory, hippocampus, fornix, thalamus, temporal cortex

This target article describes how medial temporal lobe - medial diencephalic interactions contribute to episodic memory. Previous models have focussed on neural circuitry within the temporal lobe. This earlier focus on "temporal lobe memory systems" arose from a number of assumptions about amnesia and models for amnesia. In this review these assumptions are questioned, and from this emerges a different way of considering the neural substrates of episodic memory. At the centre of this revision is the notion that the link from the hippocampus to the mamillary bodies and anterior thalamic nuclei, via the fornix, is critical for normal episodic memory (Gaffan, 1992a). Moreover, damage to this axis is responsible for the core deficits in anterograde amnesia, as originally proposed by Delay and Brion (1969). To understand why this view became unpopular and why it has now re-emerged it is necessary to describe how a number of past findings have been interpreted.

We will first describe the main features of the proposed model. This is followed by a section summarising relevant evidence from studies of amnesia and animal models of amnesia, describing the way this evidence has often been interpreted. The third section examines certain assumptions underlying previous interpretations, and shows that existing evidence can be re-interpreted in a different way. The fourth section describes new evidence from behavioural studies, human clinical studies, single unit recording studies, and brain activation studies that provide further support for the proposed model of medial temporal - medial diencephalic interactions. The final sections consider some of the implications of the model. Throughout this review we have drawn on evidence from studies of animals when the clinical data lacks sufficient anatomical resolution. Great care is needed when transposing results across species (Tulving & Markowitsch 1994) and biases can be introduced by the reliance on one particular research method. For these reasons we have tried, wherever possible, to present complementary data from an array of techniques and from more than one species.

1. Main features of proposed model:

1. The anatomical focus of the model concerns the connections between the hippocampus, the mamillary bodies, and the medial thalamus. (As a matter of terminology the term `hippocampus' is used here to refer to the hippocampal fields CA1 - 4, the dentate gyrus, and the subicular complex. The mamillary bodies and the medial thalamus are both medial components of the diencephalon, which is composed of the thalamus, hypothalamus, epithalamus and subthalamus). The hippocampal efferents to the medial diencephalon are regarded as vital for normal hippocampal activity and are, hence, seen as functional extensions of the hippocampus.

   

   Figure 1. Schematic diagram of the principal pathways that allow the encoding of episodic information and underlie recollective aspects of recognition. The relative thickness of the lines indicates the putative importance of the various connections.

   The principal thalamic targets in this system are the anterior thalamic nuclei. These nuclei receive direct hippocampal projections via the fornix, and indirect hippocampal projections via the mamillary bodies and the mamillothalamic tract. Other thalamic nuclei that may contribute to this system are the rostral midline nuclei and the lateral dorsal nucleus.

2. The system beyond the anterior thalamic nuclei becomes more diffuse, but one component projects back from the anterior thalamic nuclei to the hippocampus and to adjacent temporal cortical regions. These return connections, which mainly use the cingulum bundle, form part of a circuit that permits these diencephalic regions to influence temporal lobe processing. Other important outputs are to the cingulate and prefrontal cortices. A consequence of the diffuseness of the system beyond the anterior thalamic nuclei is that damage in the relevant tracts or regions (e.g. cingulum bundle and prefrontal cortex) has less impact upon episodic memory.
3. This extended hippocampal - diencephalic system is critical for the efficient encoding and, hence, normal recall of new episodic information. As a consequence, damage to the component structures can result in anterograde amnesia, so that a common feature of all diencephalic and temporal lobe amnesias is the bilateral involvement of part of this `extended hippocampal - diencephalic system' (i.e. the hippocampus, fornix, mamillary bodies, anterior thalamus and, possibly, the cingulum bundle). Furthermore, damage to different parts of this system produces similar memory impairments.
4. In contrast, this extended hippocampal - diencephalic system need not be vital for efficient recognition. This is because recognition is regarded as being composed of at least two independent processes (Mandler 1980), only one of which is hippocampally dependent. Thus item recognition occurs through recollection of the stimulus (`remembering'), a process which is hippocampal dependent, and by detecting stimulus familiarity (`knowing') which does not require the hippocampus. The latter process is especially dependent on the perirhinal cortex in the temporal lobes.
5. Although the hippocampus and perirhinal cortex are anatomically linked, they are not necessarily dependent on each other for their respective roles in the encoding of episodic information and familiarity based recognition. In particular, the hippocampus and perirhinal cortex each have independent links with other association cortical areas.
6. While the hippocampus is closely linked to the anterior thalamic nuclei, the perirhinal cortex is connected with the medial dorsal thalamic nucleus. These two, parallel temporal - thalamic systems have qualitatively different contributions to learning and memory. The entorhinal cortex has attributes of both systems.

   

   Figure 2. Schematic diagram of the principal pathways underlying the detection of item familiarity. The relative thickness of the lines indicates the putative importance of the various connections.

7. The traditional distinction between temporal lobe and diencephalic amnesics is misleading as both groups have damage to the same functional system. Nevertheless, the large majority of amnesics have additional pathology in certain subcortical and cortical sites and this can extend the nature of the memory loss so that it involves other aspects of memory.
8. The proposed hippocampal - diencephalic system is required for the encoding of episodic information, permitting it to be set in its spatial and temporal context (`episode'), so aiding subsequent retrieval and reducing interference (i.e. heightening discriminability).
9. The prefrontal cortex interacts with both of these systems at a variety of levels, so engaging efficient encoding strategies that can then aid subsequent recall.

2. Studies of amnesia and animal models of amnesia, and their interpretation:

2.1 Neuropathological evidence:

Anterograde amnesia is typified by a failure to acquire or retain `episodic' information (Tulving 1983) that occurred after the onset of brain injury. Damage in more than one brain region can result in anterograde amnesia, and neuropathological studies have repeatedly highlighted the medial temporal lobes and the medial diencephalon. Identifying the critical structures has, however, proved to be surprisingly difficult. Although it is often assumed that temporal lobe amnesia is principally a consequence of damage to the hippocampus, it remains to be confirmed whether such damage is sufficient to induce amnesia. Relevant evidence has come from amnesic cases with discrete unilateral hippocampal damage in one hemisphere combined with more extensive temporal lobe damage in the other hemisphere (Penfield & Mathieson 1974; Woods et al. 1982). If bilateral damage is required to induce amnesia such cases strongly implicate the hippocampus. Other evidence has come from amnesics with confirmed bilateral pathology restricted to the hippocampus and the adjacent parahippocampal gyrus or uncus (DeJong et al. 1969, Glees & Griffiths 1952). Some of the most convincing evidence has, however, come from the discovery that hypoxia can produce both a permanent anterograde amnesia and discrete bilateral hippocampal pathology (Cummings et al. 1984; Rempel-Clower et al. 1996; Victor & Agamonolis 1990; Zola-Morgan et al. 1986). There is, however, debate over whether these cases suffer `hidden' pathology (see Section 3.1), so there is still a need to confirm whether discrete, bilateral hippocampal damage can induce anterograde amnesia.

Diencephalic amnesia appears even more complex as neuropathological evidence has implicated several structures, namely the mamillary bodies, the anterior thalamic nuclei, the medial dorsal thalamic nucleus, and the parataenial thalamic nucleus (Aggleton & Sahgal 1993; Clarke et al. 1994; Dusoir et al., 1990; Mair et al. 1979; Markowitsch 1982; Parkin & Leng 1993). A number of adjacent tracts (the mamillothalamic tract and the internal medullary lamina) have also been implicated (Markowitsch 1988; Savage et al. 1997). Unfortunately, there are still no amnesic cases with confirmed, circumscribed damage in just one of these structures. Furthermore, the proximity of these nuclei to one another, along with the likelihood of damage to fibres of passage and adjacent tracts, makes it extremely unlikely that unambiguous cases will be discovered.

2.2 Testing recognition to assess anterograde amnesia in animals:

The lack of unambiguous clinical evidence has led researchers to model anterograde amnesia in animals, and so test unusually selective lesions. A prerequisite is, however, the development of behavioural tasks that tax the same classes of memory that are lost in amnesia. This need is underlined by the many examples of spared learning abilities in amnesia, which include classical conditioning, visuo-motor skill tasks, and priming (Parkin & Leng 1993; Schacter et al. 1993; Weiskrantz 1990).

Studies with animals have, in fact, relied very heavily on behavioural tests of recognition. This is because a loss of recognition is a striking feature of anterograde amnesia, and has been regarded as a core deficit (Haist et al. 1992; Parkin & Leng 1993; Squire & Knowlton, 1995; Squire & Shimamura 1986). Furthermore, the use of forced-choice designs makes it relatively easy to test animals. In contrast, examining the recall of episodic information by animals has proved much more problematic. As a consequence the favoured test of recognition, delayed nonmatching-to-sample (DNMS), has become the litmus test for models of anterograde amnesia.

In DNMS the animal is first shown a sample stimulus (often a `junk' object). After a delay the animal is then shown that same object along with a novel or less familiar object. Selection of the novel object (nonmatching) is rewarded in DNMS, while in delayed matching-to-sample (DMS) selection of the familiar object is rewarded. In the `trial-unique' version of DNMS and DMS both the novel and familiar objects are then discarded so that new items can be used for the next sample and the next novel alternative. Early studies using the trial-unique version of the DNMS task with monkeys soon confirmed that, as in people, large medial temporal lesions (Mishkin 1978; Zola-Morgan et al. 1982) and large medial diencephalic lesions (Aggleton & Mishkin 1983a; 1983b) produce very severe recognition deficits. The apparent validity of these recognition tests was further strengthened by studies showing that people with either temporal lobe or diencephalic amnesia are markedly impaired on forced-choice recognition tasks designed to be analogous to the DNMS and DMS tasks given to monkeys (Aggleton et al. 1988; Squire et al. 1988). It is therefore not surprising that these tasks have been used to assess the effects of selective bilateral damage in a number of key sites.

2.3 Testing the contribution of the fornix:

One site of especial interest has been the fornix. Among its components this tract contains the cholinergic innervation to the hippocampus from the medial septum, as well as hippocampal efferents to the diencephalon, striatum, and prefrontal cortex. These efferents include dense projections to the mamillary bodies and the anterior thalamic nuclei, which in monkeys are conveyed solely in the fornix (Aggleton et al. 1986a; Aggleton & Saunders 1997). As a consequence, the fornix forms a vital bridge between medial temporal and medial diencephalic regions implicated in anterograde amnesia.

Although the first study to use the DMS task to assess the effects of fornix transection reported an impairment (Gaffan 1974), a series of later DNMS and DMS studies found that fornix transection produced little or no recognition deficit in monkeys (Bachevalier et al. 1985a; 1985b; Gaffan et al. 1984; Zola-Morgan et al. 1989a), and spared DNMS performance in rats (Aggleton et al. 1990; Rothblat & Kromer 1991; Shaw & Aggleton 1993). Similarly, fornix lesions were found to have no effect on spontaneous tests of object recognition (Ennaceur & Aggleton, 1994; Ennaceur et al. 1996, 1997). Indeed in one study, monkeys with fornix lesions eventually performed the DNMS task significantly better than control animals (Zola-Morgan et al. 1989a), while in another study monkeys with fornix lesions showed enhanced preference for perceptual novelty (Zola-Morgan et al. 1983). Similarly, rats with fornix lesions were able to acquire a DNMS task more rapidly than control animals (Shaw & Aggleton, 1993). The immediate conclusion was that fornix damage did not disrupt recognition and, hence, was not sufficient to induce anterograde amnesia (Squire & Zola-Morgan 1991; Zola-Morgan et al. 1989a).

This conclusion was consistent with a review of memory loss and fornix damage in humans (Garcia-Bengochea & Friedman, 1987). Of 142 patients thought to have bilateral fornicotomy for the treatment of epilepsy, none had persistent memory problems. A further 13 cases with fornix damage associated with third ventricle colloid cysts were also considered. Four of these cases had persistent memory loss (Carmel 1985; Garcia-Bengochea & Friedman 1987; Sweet et al. 1959), but the likelihood that the cysts had caused additional diencephalic damage weakened the value of these individual cases. Similar constraints can be applied to other cases in whom surgery for cysts or tumours resulted in both fornix damage and memory loss (Cameron & Archibold 1981; Geffen et al. 1980; Heilman & Sypert 1977; Tucker et al., 1988). Additional problems of interpretation arise in those patients in whom the hippocampal commissures as well as the fornix were cut or disconnected (Heilman & Sypert, 1977; Tucker et al., 1988). Although Hassler (1962) described a woman in whom stereotaxic coagulation of the fornices led to an amnesic state, the woman only survived a few days after surgery so severely limiting assessment. Taken together, the cases with presumed fornix damage and apparently unchanged memory (Gargcia-Bengochea & Friedman 1987; see also Woolsey & Nelson, 1975) far outnumbered the few single case studies in which fornix damage appeared to be associated with amnesia.

Other evidence has come from studies on the mamillary bodies, which the fornix innervates. It had long been appreciated that mamillary body degeneration is a consistent feature of Korsakoff's disease and that it might contribute to the anterograde amnesia. A comprehensive neuropathological study by Victor and his co-workers (1971) concluded, however, that thalamic damage (and in particular damage to the medial dorsal thalamic nucleus) was a better predictor of the memory loss. Consistent with this was the finding that mamillary body lesions in animals did not disrupt DNMS performance (Aggleton & Mishkin 1985; Aggleton et al. 1990; Zola-Morgan et al. 1989a), while lesions in the medial dorsal thalamic region impaired both the acquisition and performance of the DMS and DNMS tasks (Aggleton & Mishkin 1983b; Hunt & Aggleton, 1991; Mumby et al., 1993; Parker et al. 1997; Zola-Morgan & Squire 1985a). As these findings failed to support a role for the mamillary bodies in anterograde amnesia they accorded with similar evidence for the fornix.

2.3 Comparing the effects of lesions in the hippocampus and lesions in adjacent cortical regions:

Studies with animals also provided the opportunity to examine the effects of increasingly selective lesions within the temporal lobe. Aspiration lesions of the hippocampus consistently produced a modest, but significant, DNMS deficit (Murray & Mishkin 1986; Zola-Morgan & Squire 1986; Zola-Morgan et al. 1989a; Zola-Morgan et al. 1993), so supporting the contribution of this structure to amnesia. More discrete temporal lobe lesions also revealed that the amygdala was not critical (O'Boyle et al. 1993; Zola-Morgan et al. 1989b). Much more surprising was the discovery that the cortex immediately lateral to the amygdala and hippocampus is of vital importance for DNMS performance. Thus lesions involving the rhinal region (comprising the perirhinal and entorhinal cortices) or more extensive lesions involving the rhinal region and the parahippocampal gyrus produce extremely severe and persistent DNMS deficits (Meunier et al. 1993; Mumby & Pinel, 1994; Murray 1996; Murray & Mishkin, 1986; Suzuki et al. 1993; Zola-Morgan et al. 1989b). More discrete lesions within the rhinal region have since highlighted the especial importance of the perirhinal cortex (Meunier et al. 1993; Meunier et al. 1996). In contrast, entorhinal lesions produce only a very mild or transient impairment (Leonard et al. 1995; Meunier et al. 1993). Similarly, removal of parahippocampal cortex does not contribute to the DNMS deficit (Meunier et al. 1996; Ramus et al. 1994). These findings, along with those from single unit recording studies (See Section 4.3), have forced a fundamental reappraisal of the contribution of individual temporal lobe structures to memory (Murray 1996).

Anatomical studies have shown that the perirhinal and parahippocampal cortices project densely upon the entorhinal cortex and, in fact, they provide nearly two thirds of the cortical inputs to the entorhinal cortex (Insausti et al. 1987; Suzuki & Amaral 1994). The entorhinal cortex itself is the major source of afferents to the hippocampus. As a consequence these indirect connections, along with a number of direct perirhinal - hippocampal projections (Suzuki & Amaral 1990; Witter & Amaral, 1991), ensure that the perirhinal and parahippocampal cortical areas are a major source of hippocampal inputs. In addition, the hippocampus has extensive reciprocal connections with the entorhinal, perirhinal and parahippocampal cortices (Suzuki 1996a; Suzuki & Amaral 1994; Witter et al. 1989). These interconnections help to reinforce the view that the hippocampus along with the perirhinal, parahippocampal, and entorhinal cortices function as a closely integrated unit subserving aspects of memory, including recognition (Squire & Zola-Morgan 1991). It should be emphasized, however, that the DNMS deficit following perirhinal removal is not simply due to a disconnection of hippocampal inputs as the severity of this deficit is considerably greater than that found after hippocampectomy (Meunier et al., 1996; Murray 1996; Zola-Morgan et al. 1993). Thus the perirhinal region must have independent mnemonic capabilities.

2.4 The temporal lobes and episodic memory: current models:

These new findings have been integrated with growing clinical evidence suggesting that hippocampal damage is sufficient to induce amnesia, and they have led to a number of influential models of temporal lobe function. A common feature of these models is that the perirhinal, entorhinal and parahippocampal cortices along with the hippocampus, form the key components of a closely integrated temporal lobe memory system. This system is reciprocal as the plentiful projections back from the hippocampus to the entorhinal cortex and the perirhinal/parahippocampal cortices are seen as instrumental in setting up long term representations (i.e. memories) in neocortex (Eichenbaum et al. 1994; Squire & Knowlton 1995; Squire and Zola-Morgan 1991; Suzuki 1996a, 1996b). One important consequence of the reciprocal nature of these interactions is that the proposed systems are largely self-contained within the temporal lobes. This has served to distance other structures such as the fornix, anterior thalamic nuclei, and mamillary bodies, and imply that the involvement of these regions in diencephalic amnesia will reflect a qualitatively different syndrome.

In one of the most cited models (Squire & Zola-Morgan 1991) the parahippocampal, perirhinal, and entorhinal cortices form a reciprocal network with the hippocampus to create a `medial temporal memory system'. This system is crucial for the rapid acquisition of new information about facts and events, which then gradually become consolidated in the neocortex and eventually become independent of the hippocampus (Squire & Alvarez 1995; Squire & Zola-Morgan 1991). The role of the hippocampus is to bind together different components of the memory. Later enlargements of this model have acknowledged some linkage with medial thalamic regions, but no apparent role has been provided for hippocampal outputs to the mamillary bodies and anterior thalamus via the fornix (Squire & Knowlton 1995; Zola-Morgan & Squire, 1993). This exclusion stems from the failure of either fornix or mamillary body lesions to disrupt DNMS performance, and the assumption that there is a close relationship between recognition and recall (Haist et al. 1992; Squire & Knowlton 1995). It is therefore presumed that these connections are not necessary for the recall of episodic (declarative) memory.

A related model (Eichenbaum et al. 1994) proposes a `hippocampal memory system' formed by the hippocampus and the `parahippocampal region' (comprising the entorhinal, perirhinal and parahippocampal cortices). This hippocampal memory system contributes both to the temporary maintenance of memories and to the processing of a particular type of memory representation. In particular, the parahippocampal region supports intermediate term storage of individual items while the hippocampal formation is concerned with organizing memories according to relevant relationships between items, including spatial relationships (Eichenbaum et al. 1994). As this `hippocampal memory system' is seen to be critical for episodic memory, dysfunction of the system can lead to anterograde amnesia.

3. A critical examination of key assumptions underlying these models of the neural substrates of recognition and recall:

In developing these models of temporal lobe involvement in episodic memory a number of different assumptions have proved very influential. These are: (i) That hippocampal damage is sufficient to impair recognition; (ii) That testing recognition (i.e. using DNMS or DMS) provides a valid assay for the core deficits in anterograde amnesia; (iii) That hippocampal function is critically dependent on afferents from the perirhinal region. There are now, however, good grounds for questioning all of these assumptions and in doing so a quite different view of temporal lobe - diencephalic interactions emerges.

3.1. Is hippocampal damage sufficient to impair recognition?

The importance of the perirhinal cortex highlights the need to re-examine the effects of hippocampectomy upon DNMS using techniques that spare rhinal regions. Interestingly, lesions of the rat hippocampus are possible via a dorsal route that avoids the rhinal cortices. Hippocampectomies performed in this manner have little or no effect upon DNMS tests (Aggleton et al. 1986b; Duva et al. 1997; Mumby et al. 1996; Steele & Rawlins 1993). Another approach has been to induce ischaemic lesions which can produce seemingly selective pathology in the hippocampus. Such lesions are accompanied by persistent DNMS deficits in both monkeys (Bachevalier & Mishkin 1989; Zola-Morgan et al. 1992) and rats (Wood & Phillips 1991; Wood et al. 1993). A problem is that the neural dysfunction caused by the ischaemia may be much more extensive than the region of gross pathology (Bachevalier & Meunier 1996; Gaffan & Lim 1991; Nunn & Hodges 1994). Occlusion of the posterior cerebral artery results, for example, in a DNMS deficit greater than that expected from the grossly apparent brain damage (Bachevalier & Mishkin 1989). Similarly, discrete ischaemic hippocampal lesions in rats produce marked DNMS deficits (Wood & Phillips 1991; Wood et al. 1993), yet neurotoxic lesions intended to match the extent of the apparent ischaemic damage have no effect on DNMS performance (Duva et al. 1997). Extensive conventional hippocampal lesions not only spare DNMS performance (Mumby et al. 1996; Wood et al. 1993) but, remarkably, can attenuate the effects of ischaemia (Mumby et al. 1996). This result not only highlights the mismatch between the observed pathology and the functional pathology, but also indicates that the ischaemia resulted in extrahippocampal dysfunctions subsequent upon the initial hippocampal pathology (Mumby et al. 1996). Finally, a recent PET study (Markowitsch et al. 1997) has highlighted the limitation of relying on MRI to uncover functional damage in cases of anoxia. This is because PET revealed widespread regions of hypoactivity in an amnesic case that could not be predicted from MRI scans (Markowitsch et al. 1997).

The possibility that ischaemia can lead to more extensive dysfunction than that apparent by standard pathological measures has, however, been disputed (Squire & Zola 1996). It has been argued that the DNMS deficits following posterior artery occlusion (Bachevalier & Mishkin 1989) were exagerated by reference to unusually high scoring controls, and that monkeys with hippocampal lesions produced by sterotaxy (Alvarez et al. 1995) perform at a comparable level to those with ischaemic lesions (Squire & Zola 1996). The first of these points requires additional control data to resolve. The second criticism is, however, potentially misleading as the comparison included data from other tests i.e. those not testing recognition. When the data are taken only from comparable DNMS tests (delays 15s to 10 mins) it is found that three of the four ischaemic monkeys performed at more than 2.7 standard deviations below the mean score of the stereotaxic hippocampectomy cases (Squire & Zola 1996), while the control animals for the two studies performed at equivalent levels. The DNMS scores of the ischaemic animals were, however, comparable to those of monkeys with hippocampal lesions made by techniques that also damage adjacent perirhinal cortex (Bachevalier & Meunier 1996). While the balance of evidence indicates that anoxia can produce more extensive recognition dysfunctions than that predicted from an assessment with standard histological methods, it is also clear that this key issue requires further examination (Nunn & Hodges 1994).

For these reasons it is preferable to focus on studies that have examined selective, stereotaxic lesions within the hippocampus. In one of the few such studies, radiofrequency lesions were placed bilaterally within the hippocampus (Alvarez et al. 1995). The lesions did not disrupt DNMS performance significantly until there was a delay of 10 min between sample presentation and test (Alvarez et al. 1995). In a number of other sterotaxic studies a neurotoxin (ibotenic acid) was injected into the monkey hippocampus, so sparing fibres of passage and adjacent fibre tracts (Beason-Held et al. 1993; Murray 1996; Murray & Mishkin 1996; O'Boyle et al. 1993). While the first of these studies reported DNMS deficits (Beason-Held et al., 1993), the remaining studies observed normal levels of performance even though the hippocampal fields CA1-4, along with the amygdala, were destroyed. In one of these studies the retention interval was extended to 40 mins, but unlike an earlier study that had found an impairment with such delays (Alvarez et al. 1995), the animals were not removed from the apparatus during testing (Murray & Mishkin 1996). These animals showed no DNMS impairment (Murray & Mishkin 1996). It therefore appears that selective hippocampal lesions can often spare DNMS performance, although for some of these reports the histology remains to be published in a comprehensive form. It is also still necessary to examine the performance of monkeys with neurotoxic hippocampal lesions that include the subiculum.

The effects of these selective hippocampal lesions now closely correspond to the effects of fornix lesions on DNMS i.e. they typically have little or no effect. This is noteworthy as fornix transection often mimics hippocampal dysfunction, most obviously for tests of spatial memory (Aggleton et al. 1986b; 1992; 1995a; Barnes 1988; Olton et al. 1982; Saunders & Weiskrantz, 1989). It had appeared that DNMS presented an important exception to this general rule, but these recent stereotaxic studies show that the effects of hippocampectomy and fornicotomy are in accord for DNMS as well.

It has been argued that the lack of a clear hippocampal lesion deficit in DNMS tasks may be due to the training prior to surgery, which is then able to mask any subsequent lesion deficit (Alvarez et al., 1995). As learning the nonmatching rule per se cannot help the animal solve any individual problem it is difficult to see how greater training could obscure a deficit unless there are ceiling effects. Nevertheless, this claim has led to a number of studies of spontaneous recognition based upon preferential viewing of novel visual stimuli. Using such tasks it has been reported that lesions not only of the perirhinal cortex (Clark et al., 1997) but also of the hippocampus (Clark et al., 1996) can disrupt performance at delays as short as 10s. Such tasks often use complex visual stimuli and previous lesion studies have demonstrated that hippocampal system lesions impair the ability to use `scenes' that are composed of an array of different features (Gaffan 1994b). Thus the abnormal behaviour following hippocampal lesons may reflect a failure to associate the component elements. It is also the case that spontaneous tests of recognition are more prone to disruption by other factors such as hyperactivity or increased distractability. In an ingenious variant on such tasks, Honey et al. (1998) showed that neurotoxic lesions of the rat hippocampus do not affect orientation and subsequent habituation to novel visual and auditory stimuli. It was, however, found that animals with these lesions failed to orient when familiar combinations of these crosmodal stimuli were rearranged (mismatched). Thus the hippocampal lesions spared novelty detection per se, but the mismatch condition revealed a failure to detect or respond to changes in the learnt association between the pairs of crossmodal stimuli (Honey et al., 1998).

The evidence showing that extensive, but selective, hippocampal damage can often spare DNMS raises the question of whether there is comparable, clinical evidence. One source of potential evidence comes from amnesic people with hypoxic damage who are very likely to suffer hippocampal damage (but may also suffer `hidden pathology' - see above). Such amnesics can show apparently normal recognition performance in spite of impaired recall (Volpe et al. 1986). Consistent with this are the findings from a recent survey of amnesics (Aggleton & Shaw 1996) which had analysed results from a standard test of recognition, the Warrington Recognition Memory Test (RMT). The RMT (Warrington 1984) consists of two subtests, one testing face recognition the other testing word recognition. From a sample of 112 amnesics placed in eleven distinct pathological groupings, it was found that three groups of amnesics failed to differ from their age-matched norms (Aggleton & Shaw 1996). One of these groups comprised patients with restricted hippocampal damage following hypoxia, while another contained patients with fornix damage (Aggleton & Shaw 1996; see also McMackin et al. 1995), and a third had selective diencephalic damage. These groups not only failed to differ from the normal subjects, they also performed significantly better than some of the other amnesic groups.

While these RMT results closely match the findings for DNMS performance by nonhuman primates i.e. little or no effect following hippocampal or fornix damage, there are a number of important constraints. The first is that the RMT data come from just one test of recognition and, as was indicated in the first section of this review, it is to be predicted that hippocampal damage will have more impact on some tests of recognition than others. The second is that cases with anoxic damage may have variable covert pathology. Both of these considerations apply to a recent review of recognition following anoxic hippocampal damage in humans (Reed & Squire 1997), which convincingly shows that this aetiology can lead to recognition deficits across a wide range of tests. Even so, when compared to test norms, performance on the standard version of the RMT is apparently preserved in some of these cases and deficient in others (Reed & Squire, 1997). A related case concerns an amnesic who performed very poorly on the RMT, even though magnetic resonance imaging (MRI) studies indicated that the subject had circumscribed lesions confined to areas CA1 and CA2 (Kartsounis et al. 1995). This same person did, however, show very severe retrograde amnesia suggestive of more extensive cortical damage (Kapur et al. 1992; Zola-Morgan et al. 1986). In view of the fact that the amnesia arose from repeated ischaemic episodes, this apparent discrepancy may relate to the issue of hidden pathology (Bachevalier & Mishkin 1989; Mumby et al. 1996).

Other relevant evidence comes from a recent study of 104 epileptic patients who had been tested on the RMT and had unilateral temporal lobe pathology confirmed by MRI (Baxendale 1997). Patients with combined cortical and hippocampal damage performed significantly worse than those with selective hippocampal damage. Furthermore, the group mean score of those with selective left hippocampal damage on the test of word recognition (the subtest they should be most impaired on) was in the normal range, as was the group mean score for those with right hippocampal damage on the face recognition test (Baxendale 1997). The conclusion, that unilateral hippocampal damage had no consistent effect on this test of recognition, reinforced a previous study showing that hippocampal sclerosis had no apparent effect on either of the RMT subtests (tested preoperatively), although deficits on delayed recall were found (Miller et al. 1993). These conclusions appear to contrast with a recent study using event related potentials which showed a loss of reactivity to novel stimuli in five subjects with combined unilateral pathology in the hippocampus (Knight 1996). In all five cases, however, the pathology involved the parahippocampal gyrus and the entorhinal cortex (Knight 1996), so the resulting deficit can be predicted.

Even if it is accepted that hippocampectomy can induce a subtle DNMS deficit (Alvarez et al. 1995; but see Murray & Mishkin 1996), this is only apparent after very lengthy delays e.g. 10 min. This contrasts with amnesic subjects who are typically impaired on DNMS and DMS tasks after delays of only 40s between sample presentation and test (Aggleton et al. 1988; Holdstock et al. 1995; Squire et al. 1988). Furthermore, amnesic subjects show significantly faster rates of forgetting over these relatively short delays (Holdstock et al. 1988), while monkeys with selective hippocampectomy do not. These differences suggest that hippocampal damage in monkeys is not sufficient to reproduce the recognition deficit typically found in amnesia.

A final factor concerns the type of stimulus being tested. Studies with rats have shown that both fornix transection and hippocampectomy can disrupt recognition when large, relatively featureless stimuli (test boxes) are used instead of trial-unique discrete objects (Cassaday & Rawlins 1995; Rawlins et al. 1993). This impairment is most evident when the plain boxes are used repeatedly within a session i.e. are not trial unique (Rawlins et al. 1993), but deficits are also observed when discrete objects are placed in these large test boxes (Cassaday & Rawlins 1997). A plausible explanation of these results is that the animal encodes the large box or the stimuli inside the large box as part of a spatial (scenic) array rather than as a discrete stimulus (Cassaday & Rawlins 1995), thus rendering it sensitive to hippocampal dysfunction. When the boxes are small they are encoded as objects and no deficit is seen (Cassaday & Rawlins 1997). Similarly, studies with monkeys have shown that fornix lesions can reliably disrupt the recognition of `scenes' in which common elements are repeated but occur in different spatial configurations (Gaffan 1991; 1992b; 1994b). These impairments can be directly related to the widely accepted view that the hippocampus is vital for the efficient encoding of allocentric space (O'Keefe & Nadel 1978). The importance of stimulus type is further emphasized by the spontaneous orientation task by Honey et al. (1998) and by recent activation studies (see Section 4.4).

3.2. Does testing recognition provides a valid assay for anterograde amnesia?

A closely related debate is whether tests such as DNMS and DMS are a valid assay for amnesia. One view is that recognition is an integral part of declarative memory (Haist et al. 1992; Knowlton & Squire 1995) as people can subjectively evaluate their memory and either retrieve items (recall) or make judgements as to their previous occurrence (recognition). This model tightly links the two processes and so predicts that anterograde amnesia will impair both recall and recognition, and that the deficits will be related. An alternate view is that recognition and recall depend, in part, on different processes. One such view is that recognition benefits from an additional component of processing that is based on `perceptual fluency' or `feelings of familiarity' (Gardiner 1988; Gardiner & Parkin 1990; Jacoby 1991; Mandler 1980; see Section 5). This process is regarded as being additive and separate to the explicit memory of an event (Mandler 1980), and corresponds to feelings of `knowing' that something is familiar rather than `remembering' (i.e. recalling) its previous occurrence (Gardiner 1988; Gardiner & Parkin 1990). As a consequence it may be predicted that a loss of episodic memory need not always be accompanied by a comparable loss of recognition.

In fact, a number of reports have described individual amnesic cases (Dusoir et al. 1990; Gaffan et al. 1991; Hanley et al. 1994; Parkin & Hunkin 1993; Parkin et al. 1993) or even groups of amnesics (McMackin et al. 1995; Volpe et al. 1986) with relatively preserved recognition. For example, a group of subjects with bilateral fornix damage following third ventricular cysts (McMackin et al. 1995) were able to perform the RMT tasks within normal limits, even though they were clearly impaired on tests of episodic memory. Individual cases of interest include a man who suffered bilateral traumatic injury to the mamillary body region (Dusoir et al. 1990), in whom PET studies revealed additional hypoactivity in the left hippocampus (Kapur 1995). A clear and persistent anterograde amnesia developed, yet he performed well within normal limits on a series of recognition tests including the RMT (Dusoir et al. 1990). He also performed very well on a DMS task using single abstract patterns (Holdstock et al. 1995).

Figure 3. Performance of an amnesic subject with bilateral mamillary body damage (BJJ) showing spared performance compared to normal controls and a group of mixed amnesics on a delayed matching-to-sample task using visual patterns (data from Holdstock et al. 1995).

This is of interest as the task avoided ceiling effects even though it used a DMS procedure to assess the retention and recognition of single stimuli. Furthermore, all of the other amnesic subjects tested on the same DMS task were markedly impaired, even though their delayed recall deficits (as measured by the Wechsler Memory Scale Revised - WMSr) were comparable to those of the mamillary body case (Holdstock et al. 1995).

Other individual cases include a person who had suffered a hypothalamic tumour close to the mamillary bodies, and who displayed a severe anterograde amnesia (Parkin & Hunkin 1993). This patient achieved scores in the 83rd (words) and 94th (faces) percentile on the two RMT subtests (in stark contrast to a score of 56 on the WMSr Delayed Recall index). Of similar interest was a young woman who displayed Wernicke's encephalopathy following a relatively brief history of alcoholism (Parkin et al. 1993). She showed a chronic, profound impairment on tests of recall but her recognition memory was remarkably well preserved across a variety of tests. These included the RMT on which she scored in the 75th percentile for both words and faces (Parkin et al. 1993), as well as normal performance on a more difficult RMT variant in which the face stimuli were presented upside down.

These examples of spared recognition do not simply arise because tests of recognition are easier to perform than tests of recall, or because the individual cases only suffered from very mild amnesic syndromes. The former can be excluded as a number of studies have taken special care to preclude ceiling effects (Hanley et al. 1994; Holdstock et al. 1995; Parkin et al. 1993; Volpe et al. 1986). Similarly, differences in the severity of the anterograde amnesia can also be discounted as performance on other memory tests has been carefully documented in individual cases with spared recognition (Aggleton & Shaw 1996; Dusoir et al. 1990; Hanley et al. 1994; Parkin & Hunkin 1993; Parkin et al. 1993).

3.3 Is hippocampal function dependent on afferents from the perirhinal region?

One of the more surprising aspects of the current model is the supposed extent to which some hippocampal functions are independent of their perirhinal inputs. This appears surprising for at least two reasons. First, there are many direct and indirect connections between the two regions. Indeed, the perirhinal and parahippocampal cortices combined provide approximately two thirds of the inputs to the entorhinal cortex (Insausti & Amaral 1987; Suzuki 1996a; Suzuki & Amaral 1994), which is the cortical gateway to the hippocampus. In addition, the hippocampus projects directly upon the perirhinal cortex and entorhinal cortex; the latter projects, in turn, to the perirhinal and parahippocampal cortices (Saunders & Rosene 1988; Suzuki 1996; Suzuki & Amaral 1994). Second, the type of information that appears to be present in the perirhinal cortex (see Section 4.3) could provide the elemental fragments that an episodic `memory system' might operate upon (Brown 1990; Eichenbaum et al. 1994; Gaffan & Parker 1996; Squire & Zola-Morgan 1991). Not surprisingly, both classes of evidence feature in previous models of medial temporal lobe function (Eichenbaum et al. 1994; Squire & Zola-Morgan 1991). As a consequence these models predict that perirhinal damage should disconnect the hippocampus and so mimic the effects of hippocampal removal. For this reason it should not be possible to produce a double dissociation between these two regions.

Recent studies on the perirhinal cortex do, however, suggest that this cortical region has a different relationship with the hippocampus to that proposed by previous models (Eichenbaum et al. 1994; Squire & Zola-Morgan 1991). Most striking is lesion evidence showing that functions of the hippocampus and the perirhinal cortex can be doubly dissociated from one another. In one study fornix lesions in monkeys produced severe deficits on a spatial discrimination and reversal task (Gaffan, 1994a). This accords with previous studies showing the sensitivity of this spatial task to lesions in the hippocampus and mamillary bodies, as well as the fornix (Aggleton & Mishkin 1985; Jones & Mishkin 1972; Mahut 1971, 1972). In contrast, perirhinal lesions had no apparent effect on the same spatial task (Gaffan, 1994a). The same study also tested recognition for visual scenes and found that on this task the perirhinal lesions produced a severe deficit while the fornix lesions resulted in a much milder impairment (Gaffan, 1994a). This double dissociation shows that the perirhinal cortex is not a critical way-station for all hippocampal inputs, and suggests that the mnemonic contributions of the two regions can differ substantially.

Evidence for a similar double dissociation has recently been uncovered in rats (Ennaceur et al. 1996; Ennaceur & Aggleton 1997). While fornix lesions severely impaired tests of spatial working memory (T-maze alternation, radial arm maze nonmatching, and delayed nonmatching-to-position in a Skinner box), cytotoxic perirhinal lesions had no apparent effect (Table 1). In contrast, only the perirhinal lesions disrupted a test of object recognition (Ennaceur et al. 1996). Although they were not tested simultaneously, other experiments have shown that lesions of the hippocampus and the anterior thalamic nuclei consistently disrupt these same spatial tasks (Aggleton et al. 1986b; 1995a; 1996), but have no apparent effect on object recognition (Aggleton et al. 1986b; 1995a). Conversely, large neurotoxic lesions including the perirhinal cortex, postrhinal cortex and area TE impaired object recognition but had no apparent effect on T-maze alternation (Aggleton et al. 1997). This last result is especially surprising as the postrhinal cortex offers an alternative route for spatial information to reach the hippocampus (Naber et al. 1997). Additional evidence for a double dissociation between hippocampal and perirhinal functioning has come from a series of c-fos activation studies (see Section 4.5). These indicate that exposure to novel visual stimuli increases neuronal activity in the perirhinal cortices but not in the hippocampus (Zhu et al. 1995b; 1996). In contrast, exposure to a novel environment can raise hippocampal activity but not perirhinal activity (Zhu et al. 1997).

	HIPPOCAMPUS	FORNIX	ANTERIOR THALAMUS	MAMILLARY BODIES	CINGULUM BUNDLE	MEDIAL DORSAL TH.	PPR CORTEX
DNMP	X	X	X	Ã	Ã	Ã	Ã
T-MAZE	X	X	X	X	X	Ã	Ã
RADIAL ARM MAZE	(X)	X	X	X	X	Ã	Ã
SWIM MAZE LATENCY	(X)	X	X	(X)	X	(Ã)	Ã
SPONTANEOUS OBJECT RECOGNITION	_	Ã	Ã	_	Ã	_	X
DNMS	Ã	Ã	_	Ã	_	X	(X)

Table 1. Effects of lesions in rats showing double dissociations between spatial memory (DNMP, T-Maze, Radial arm maze, Swim maze) and object recognition (spontaneous object recognition, DNMS). An `X' indicates a deficit, a `Ã ` indicates no effect, while `-` shows that the results have not been reported. All of the data are drawn from published research in the laboratory of the authors, with the exception of those in parenthesis. These are from Kolb et al. 1982, Morris et al. 1982; Mumby & Pinel 1994; Sutherland & Rodriguez 1989. Abbreviations: PPR, perirhinal/postrhinal cortices.

These results lead to the prediction that selective damage to the human homologue of the perirhinal cortex will impair some memory functions, including recognition, but not produce a full amnesia. Kapur et al. (1994) described a person with extensive bilateral damage to temporal neocortex combined with apparent sparing of the hippocampus and amygdala. This pathology spared parts of the entorhinal cortex but appeared to invade much of the perirhinal cortex (Kapur et al. 1994). In spite of some everyday memory difficulties, the patient did not suffer from anterograde amnesia (e.g. his WMSr delayed memory quotient was 99). He did, however, show a retrograde amnesia and a `semantic' memory loss. Furthermore, his recognition performance on the faces test of the RMT was severely impaired although word recognition appeared normal, suggesting a hemispheric difference in the extent of pathology.

Other evidence concerns a women who suffered bilateral damage to the rostral temporal cortex, while subcortical regions appeared intact (Kapur et al. 1992; see also Markowitsch et al. 1993). She displayed a severe retrograde amnesia but only a very mild loss of new learning. Of especial interest was the finding that her performance on the faces subtest of the RMT was impaired, yet on recall tests of visual non-verbal memory her performance was excellent (Kapur et al. 1992). Damage to the parahippocampal gyrus was the best predictor of the recognition memory deficit, but it was not associated with anterograde amnesia. Other relevant evidence comes from a description of five cases with a profound loss of semantic information associated with focal temporal lobe atrophy (Hodges et al. 1992). One of the key features of these subjects with `semantic dementia' was the relative preservation of episodic memory (Hodges et al. 1992). Another striking feature of semantic dementia is the finding that the loss of past autobiographic (episodic) information can show a reverse Ribot effect i.e. relative sparing of recent memories (Graham & Hodges 1997). This is the opposite to that observed in amnesic syndromes and in Alzheimer's disease (Graham & Hodges 1997), and hence points to dissociable functions played by the hippocampal system and its neighbouring cortices.

Further evidence for this dissociation comes from a series of three adolescents who suffered bilateral hippocampal pathology at birth or aged 4 or 9 (Vargha-Khadem et al. 1997). All three show a clear anterograde amnesia affecting episodic memory yet, remarkably, they have attained levels of language competance and factual knowledge that are within the low to average range. Not only do they show an apparent dissociation between semantic and episodic memory, they also show evidence of relatively preserved recognition (forced-choice) memory in the face of deficient spatial and temporal memory (Vargha-Khadem et al. 1997). While potentially important factors such as developmental reorganisation must be considered, these cases provide further evidence against the view of an interdependent relationship between the hippocampus and the temporal (perirhinal) cortices.

4. Recent support for the proposed model of hippocampal - diencephalic interactions

4.1 Behavioural evidence from lesion studies in animals: Spatial memory and scene memory

In recent years Aggleton and his collaborators have systematically examined the involvement of various limbic brain sites in the performance of tasks dependent on normal hippocampal function (Table 1). These experiments, which have used rats, have focussed on tests of allocentric spatial memory (O'Keefe & Nadel 1978). Studies using forced spatial alternation in a T-maze have revealed that normal performance depends on the integrity of the anterior thalamic nuclei, the mamillary bodies, and the cingulum bundle as well as the hippocampus and fornix (Aggleton et al. 1986b; 1995a; 1995b; 1996; Aggleton & Sahgal 1993; Neave et al. 1997). Furthermore, probe tests have confirmed that normal rats use allocentric cues to solve this spatial alternation task (Aggleton et al. 1996; Neave et al. 1997). Not surprisingly, lesions in these same sites (i.e. the anterior thalamic nuclei, the mamillary bodies, the cingulum bundle, and fornix) can disrupt other spatial tasks thought to tax allocentric spatial processing. These include the radial arm maze and the Morris water maze (Aggleton et al. 1996; Byatt & Dalrymple-Alford 1996; Neave et al. 1997; Sutherland & Rodriguez 1989; Warburton et al. 1997, 1998; Warburton & Aggleton 1998). These deficits are selective, however, as lesions in the same sites (i.e. fornix, anterior thalamic nuclei, mamillary bodies, and cingulum bundle) do not disrupt a comparable egocentric spatial task in which the animals are rewarded for turning in a constant direction while allocentric cues became irrelevant (Aggleton et al. 1996; Neave et al. 1997).

By using a standard alternation task it has been possible to compare the severity of the spatial deficits following various limbic lesions. These comparisons show that the alternation deficit is greatest after lesions in the hippocampus, fornix and anterior thalamic nuclei, and least after mamillary body damage or cingulum bundle damage.

Figure 4. T-maze alternation: Combined summary data from three experiments (no. 1, Aggleton et al. 1996 ; no. 2, Neave et al. 1997; no. 3, Warburton et al. 1997a) showing the effects of selective limbic lesions upon spatial alternation. The histograms show the mean percent correct scores for each of the groups over 15 acquisition sessions (90 trials). The lesion locations, going from left to right are: Sham controls; anterior thalamic nuclei plus lateral dorsal (ANTLD1); fornix (FX); anterior thalamic nuclei (ANT); bilateral cingulum bundle (CCB2); mamillary bodies (MB2); anterior ventral/anterior dorsal thalamic nuclei (AD1); anterior medial thalamic nucleus (AM1). The numbers 1, 2 and 3 refer to the number of the experiment. Although some of these comparisons are affected by floor effects they do serve to underline the importance of the anterior thalamic nuclei. Furthermore, because the anterior thalamic lesion effects are significantly greater than those observed after mamillary body lesions (Aggleton et al. 1995; Aggleton & Sahgal, 1993) the results point to a system subserving allocentric spatial memory that involves the direct hippocampal - anterior thalamic projections as well as the indirect hippocampal - mamillary body - anterior thalamic projections (see also Table 1.).

Figure 5. Schematic diagram showing the main group of interconnections underlying allocentric spatial memory performance in the rat. The thickness of the line corresponds to the relative importance of the connection as determined by lesion experiments.

This is consistent with the finding that complete or near-complete neurotoxic lesions of all three anterior thalamic nuclei produce an impairment as severe as that observed after fornix transection (Warburton et al. 1997).

This focus on the anterior thalamic nuclei raises the question of whether any of the three component nuclei (anterior ventral, anterior dorsal, anterior medial) is especially critical for spatial memory. All three nuclei have substantial connections with the hippocampus, mamillary bodies, and cingulate cortices (Shibata 1992; 1993a; 1993b), although there are some distinct differences in the detailed pattern of these connections. Most notably, the anterior dorsal nucleus receives afferents from the lateral mamillary nucleus while the anterior ventral and anterior medial nuclei receive their afferents from the medial mamillary nucleus (Cruce, 1975). The anterior dorsal nucleus also receives the fewest hippocampal inputs (Aggleton et al. 1986a). Single unit recording studies also point to differences within the anterior thalamic nuclei as the distribution of `head direction' cells varies within the anterior thalamic nuclei (Taube, 1995), but this has yet to be matched to any particular anatomical boundary or projection zone.

In order to investigate possible functional differences within the anterior thalamic nuclei the effects of lesions in the anterior medial nucleus have been contrasted with more lateral lesions involving both the anterior ventral and the anterior dorsal nuclei (Aggleton et al. 1996; see also Byatt & Dalrymple-Alford 1996). Both sets of lesions produced mild, but significant, deficits on the T-maze alternation task, but it was only when the lesions were combined that the full effect of anterior thalamic damage became evident. These results not only suggest that all three anterior thalamic nuclei are integral to the proposed system, they also show that attempts to assess fully the effects of anterior thalamic damage should involve all three nuclei. These findings may therefore help to account for those studies in which small, subtotal anterior thalamic lesions had little or no apparent effect on tests of spatial working memory (Beracochea et al. 1989; Beracochea & Jaffard 1995; Greene & Naranjo 1986). They also help to explain some of the deficits reported after lesions of the internal medullary lamina following pyrithiamine-induced thiamine deficiency (Langlais & Savage 1995). This is because damage to the anterior thalamic nuclei and mamillary bodies is a consistent feature of this animal model (Langlais & Savage 1995), and could account for many of the spatial deficits.

Damage to a number of other sites can disrupt T-maze alternation, including the prelimbic (medial prefrontal) and cingulate cortices (Brito et al. 1982; Markowska et al. 1989; Shaw & Aggleton 1993; Sutherland & Hoesing 1993; Sutherland et al. 1988; Thomas & Brito 1980). Both regions are of interest as they have connections with the hippocampus and anterior thalamic nuclei, as well as the medial dorsal nucleus of the thalamus. While both cortical regions presumably contribute to the normal processing of these spatial tasks, their importance may have been overestimated. This is because most lesion studies have damaged fibres of passage and adjacent tracts (e.g. the cingulum bundle). By using cytotoxins to produce selective lesions in these cortical areas evidence is emerging that even extensive damage to the cingulate cortices has little, if any, effect on spatial tasks such as T-maze alternation (Aggleton et al. 1995b; Neave et al. 1994) or the Morris water maze (Warburton et al. 1998). Similarly, more selective prefrontal lesions often produce only transient deficits on standard tasks thought to assess allocentric spatial memory (Aggleton et al. 1995b; Shaw & Aggleton 1993; Thomas & Brito 1980; Thomas & Spafford 1984). Furthermore, when more permanent deficits are observed after prefrontal damage they often appear qualitatively different to those observed after hippocampal damage e.g. they reflect a loss of behavioural flexibility rather than a loss of spatial memory (Aggleton et al. 1995b; Bruin et al. 1994; Granon et al. 1994).

A further diencephalic region that may prove to be of importance is the thalamic midline. Nuclei in this region, most especially the more rostral portions of nuclei such as reuniens, paraventricular, and parataenialis all have reciprocal connections with the hippocampus. These connections are found across a range of species, although they appear particularly dense in the rat. While highly selective anterior thalamic lesions show that damage to these midline nuclei is not necessary in order to induce a spatial alternation deficit (Aggleton et al. 1996), this does not show whether these nuclei contribute to the effects of more extensive lesions. Evidence of a possible involvement of the midline nuclei in the amnesia associated with paramedial thalamic infarcts has come from a PET study showing widespread cortical hypometabolism that might be attributable to the loss of midline nuclei (Levasseur et al. 1992). The contributions of these nuclei clearly require systematic investigation, though this will prove technically difficult.

The studies so far cited have concerned rats or mice, as comparable tests of allocentric spatial memory have rarely been conducted with monkeys. It has, however, been shown that fornix lesions impair T-maze forced alternation by monkeys, while cingulate gyrus lesions have little or no effect (Murray et al. 1989). Related evidence comes from studies showing that hippocampal and fornix lesions can both impair tasks that require the animal to remember the position of a given object (Parkinson et al. 1988; Gaffan & Harrison, 1989) or perform a place discrimination and subsequent reversals (Gaffan 1994a; Jones & Mishkin 1972; Mahut 1972). The effects of discrete anterior thalamic lesions have yet to be assessed in monkeys on such tasks, but the pattern of deficits following lesions in the mamillary bodies (Holmes & Butters 1983) strongly suggests that spatial deficits will be found. This overall pattern of results closely accords with findings from rats.

Other relevant evidence has come from an ingenious series of experiments showing how fornix damage in monkeys disrupts discrimination tasks in which task performance is aided by the ability to remember the spatial disposition of the elements that comprise the stimulus. Initial evidence came from the finding that fornix lesions disrupt conditional tasks that tax the ability to identify a particular place (Gaffan & Harrison 1989). Importantly, this deficit was only evident when the places to be distinguished contained common elements that were spatially rearranged (Gaffan & Harrison 1989). It was proposed that the fornix is necessary for the creation of a snapshot memory which stores the spatial arrangements of the items in a `scene' (Gaffan 1991). This conclusion was supported by the finding that lesions of the fornix will disrupt the acquisition of concurrent discriminations when the stimuli to be discriminated are complex scenes that often contain common elements (Gaffan 1992b; 1994a). This finding has recently been explored in more detail and it appears that the critical feature is whether task performance (in this case concurrent discrimination) is aided if the animals can remember the background and location of the stimuli to be discriminated (Gaffan 1994b).

These results have been taken to indicate that the fornix, and hence the hippocampus, is important for the scene-specific memory of objects (Gaffan 1992a, 1994a, 1994b). Using the same `object-in-place' task there is now evidence that the mamillary bodies are involved in this same process (Parker & Gaffan 1997a). Furthermore, this mamillary body involvement appears to be via its afferents from the fornix (Parker & Gaffan 1997a). Consistent with the main proposals in this review, this object-in-place task also depends on the integrity of the anterior thalamic nuclei in monkeys (Parker & Gaffan 1997b). These studies provide important evidence that the hippocampal - fornix - anterior thalamic system might aid the normal recall of episodic information as it permits the subject to distinguish or re-create the unique scene associated with the to-be-remembered item (Gaffan, 1992a, 1994b; Tulving, 1983), a process that will reduce interference from other similar events (Gaffan, 1994b). Interestingly, surgical ablations of the cingulate cortex produced only a mild impairment (Parker & Gaffan 1997b) and so echo the effects of selective cingulate/cingulum bundle lesions on spatial memory tasks in rats (Neave et al. 1994; 1996; 1997).

It has been pointed out that the anterior thalamic nuclei and the lateral dorsal nucleus contain `head direction' units (Mizumori & Williams 1993; Taube 1995). These inform the animal of the direction it is pointing, irrespective of actual position in space. As a loss of thalamic `head direction' information appears to disrupt `place' cells in the hippocampus (Mizumori et al. 1994), this could account for some of the similarities between the effects of damage in these two regions on tests of spatial navigation. This could also explain how anterior thalamic lesions can disrupt the Morris water maze to a greater extent than fornix lesions (Warburton & Aggleton 1998), the additional deficit reflecting the loss of head direction information. This role in providing head direction information is not, however, sufficient to explain the full array of similarities between the effects of hippocampal and anterior thalamic damage.

Perhaps most important are the findings that lesions of the mamillary bodies and anterior thalamic nuclei in monkeys disrupt the `object-in-place' discrimination task, and that this effect is dependent on inputs from the fornix (Parker & Gaffan 1997a,b). It is difficult to imagine how a loss of `head direction' units could account for a deficit in such a discrimination as the animal is pointing to the whole scene. Other evidence comes from the finding that selective lesions of the anterior medial thalamic nucleus in rats are sufficient to disrupt tasks sensitive to hippocampal dysfunction e.g. T-maze alternation or radial-arm maze (Aggleton et al. 1996; Byatt & Dalrymple-Alford 1966), even though there is no evidence that this division of the anterior thalamic nuclei contains head direction units (Taube 1995). Single-unit recordings also highlight the contribution of the anterior thalamic nuclei to discriminative avoidance tasks that have minimal spatial demands (Gabriel 1993). Finally, studies of neuronal activation in monkeys using the 2 deoxy-glucose method have revealed similar increases in activity in the hippocampus, mamillary bodies and anterior thalamic nuclei on a variety of tests of working memory (Friedman et al. 1990), even though performance on some of the tasks is unlikely to involve head direction information.

4.2. Reconsidering clinical evidence in the light of the proposed model:

A central aspect of the current proposal is that bilateral damage to either the fornix, mamillary bodies or anterior thalamic nuclei is sufficient to induce anterograde amnesia. Furthermore, the more selective the damage the greater the disparity between the loss of episodic memory and the sparing of recognition.

4.2.1. Fornix damage and amnesia:

A highly influential review on the effects of fornix section (Garcia-Bengochea & Friedman 1987) concluded that damage to this tract was not sufficient to induce amnesia. This review has, however, been strongly criticised (Gaffan & Gaffan 1991). One problem concerns the need to separate those cases with bilateral and unilateral fornix surgery. This is because unilateral hippocampal ablation in cases of epilepsy does not bring about anterograde amnesia and, hence, neither should unilateral fornix section. Gaffan & Gaffan (1991) pointed out that many of the cases regarded as having bilateral fornix sections (Garcia-Bengochea & Friedman 1987) did, in fact, only have unilateral surgery. Other problems with this material include the lack of pre and post-surgical psychometric data. This information is required as some of the relevant cases were psychotic or mentally retarded (Gaffan & Gaffan 1991; Sugita et al. 1971), making it difficult to measure any change in memory.

There are now a growing number of reports that have linked fornix damage with a loss of episodic memory. Most of these concern the outcome of cysts or tumours in the third ventricle (Calabrese et al 1995; Cameron & Archibold 1981; Gaffan et al. 1991; Geffen et al. 1980; Heilman & Sypert 1977; Hodges & Carpenter 1991; McMackin et al. 1995; Sweet et al. 1959; Tucker et al. 1988). Of especial interest is a report by McMackin et al. (1995) in which six patients who had received surgical removal of a third ventricular cyst were assessed on a variety of memory tasks. The status of the fornix was examined using MRI, and this revealed a clear association between bilateral fornix damage and a loss of recent verbal and nonverbal memory. This was found even in cases with no signs of ventricular dilation. The only case to show relatively normal performance on the verbal tests was a patient in whom the left fornix was intact (McMackin et al. 1995). This complements other evidence of an association between left fornical damage and poor performance on tests of verbal memory (Cameron & Archibold 1981; Hodges & Carpenter 1991; Tucker et al. 1988). While there exist a very small number of reports of bilateral fornix damage associated with tumours or surgery for cysts that did not appear to disturb memory (Cairns & Mosberg 1951; Woolsey & Nelson 1975) the import of these studies is limited by a lack of psychometric information.

4.2.2. Mamillary bodies and amnesia

While mamillary body pathology has repeatedly been associated with memory loss it has long been uncertain whether damage in this structure is sufficient to induce amnesia. Initial evidence came from the consistent necrotic state of the medial mamillary bodies in Korsakoff's syndrome (Delay & Brion 1969; Gudden 1896; Mair et al. 1979; Rigges & Boles 1944; Victor et al. 1971). Although this disease typically affects multiple brain sites (Victor et al. 1971) there are a small number of amnesic cases in which the pathology appeared to have been restricted to the mamillary bodies (Colmant 1965; Delay & Brion 1969; Remy 1942; Torvik 1987). Caution is required, however, as it is often unclear whether all other regions have been studied (e.g. frontal or temporal lobe areas). Nevertheless, the consistency of mamillary body damage in well characterised cases of Korsakoff's disease is striking (Mair et al. 1979; Mayes et al. 1988). Other relevant evidence comes from reports of amnesia following tumours located in the region of the mamillary bodies (Benedek & Juma 1941; Assal et al. 1976), but interpretation is hindered by the possible effects of raised intraventricular pressure on other diencephalic regions.

There are a few examples where pathology appears confined to the mamillary bodies and yet no amnesia is reported. The most influential of these negative examples are the five Wernicke cases listed by Victor et al. (1989). These are clearly very important exceptions as they are counter to the current proposals. Unfortunately there is no published psychometric data for these five cases, leaving it uncertain as to how it could be determined that their memories were intact (Victor et al. 1989). Furthermore, while Victor et al. (1989) argue that damage to the medial dorsal nucleus of the thalamus is consistently associated with amnesia, other studies have failed to find this precise linkage (Mair et al. 1979; Torvik 1987). In view of the widespread dysfunctions revealed by PET studies of Korsakoff's disease (Fazio et al. 1992; Paller et al. 1997), it is most unlikely that studies of Korsakoff's disease can resolve this issue.

Other evidence comes from the amnesia that can follow thalamic vascular lesions. A consistent feature of the pathology is damage to the mamillothalamic tracts (Castaigne et al. 1981; Cramon et al. 1985; Gentilini et al. 1987; Graf-Radford et al. 1990; Goldenberg et al. 1983; Hodges & McCarthy 1993; Mori et al. 1986; Parkin et al. 1994; Stuss et al. 1988). This pathway carries projections from the mamillary bodies to the anterior thalamic nuclei, and so is integral to the circuitry in this proposal. The thalamic pathology does, however, often invade other regions including the medial dorsal nucleus and the internal medullary lamina, both of which have also been implicated in memory dysfunction. The conclusions of two surveys that compared the extent of diencephalic pathology with the presence and severity of amnesia are therefore of particular value (Cramon et al. 1985; Gentilini et al. 1987). Both studies failed to find an association with damage to the medial dorsal thalamus (Cramon et al. 1985; Gentilini et al. 1987; see also Daum & Ackerman 1994; Graf-Radford et al. 1990; Markowitsch 1982). One of these surveys also found no consistent link with internal medullary lamina damage (Gentilini et al. 1987). In all of these studies, however, the presence of bilateral mamillothalamic tract damage was a reliable predictor of amnesia (Cramon et al. 1985; Gentilini et al. 1988).

The corollary of this conclusion is that cases in which the mamillothalamic tract and the anterior thalamic nuclei are spared will not become amnesic. This is supported by the failure of lesions restricted within the medial dorsal nucleus to produce anterograde amnesia (Kritchevsky et al. 1987; Markowitsch 1982). Exceptions to this prediction are case 5 of Castaigne et al. (1981) and a single case described by Calabrese et al. (1993). In the latter case the pathology is based on CT scans (Calabrese et al. 1993) and hence may lack sufficient resolution. In the former case the thalamic pathology is confirmed by post mortem (Castaigne et al. 1981), but the fact that this patient suffered from hypertension and loss of consciousness prior to the infarct raises the possibility of pathology in other key brain sites not reported.

Penetrating brain injuries in the region of the mamillary bodies have been described in two amnesic patients (Dusoir et al. 1990; Squire et al. 1989). One of these cases (BJ) appeared to suffer bilateral damage closely confined to the mamillary body region (Dusoir et al. 1990), and although the subject showed some recovery he has a permanent impairment for the recall of verbal and nonverbal material (Dusoir et al. 1990; Kapur et al. 1995). The severity of his amnesia, as measured by the WMSr, is typical for anterograde amnesia (Butters et al. 1988). The nature of BJ's injury does, however, mean that other diencephalic pathways could have been damaged (Kapur et al. 1995). Indeed, a PET study of B.J. revealed evidence of unilateral hippocampal hypoactivity in the same hemisphere as received the largest amount of mamillary body damage (Kapur et al. 1995). It is not, however, possible to tell if this hypoactivity reflects a primary pathology or a secondary response to the mamillary body damage. Taken together, there is still no single example of an amnesic with discrete, bilateral mamillary body damage. Nevertheless, the weight of evidence strongly indicates that damage to this region is sufficient to impair episodic memory.

4.2.3. The anterior thalamic nuclei and amnesia in humans

Direct evidence implicating the anterior thalamic nuclei in diencephalic amnesia is very limited. In Korsakoff's syndrome, damage to the anterior nuclei only occurs in about one third of all cases (Victor et al., 1971), while the lateral dorsal nucleus shows degeneration in the majority of cases (Brion & Mikol 1978; Victor et al. 1971). Although infarcts of the paramedian thalamic arteries often spare the anterior thalamic nuclei, they do consistently deafferent the anterior thalamic nuclei by disrupting the mamillothalamic tract.

More direct evidence comes from a handful of cases with anterior thalamic damage. For example, a recent MRI study described a case of anterograde amnesia associated with an infarct largely confined to the left anterior thalamic nuclei and the left mamillothalamic tract in which a loss of verbal memory was most pronounced (Clarke et al. 1994). A PET study of the same subject revealed decreased metabolic rate in the posterior cingulate cortex (Clarke et al. 1994), a result consistent with pathology in the anterior thalamic nuclei. Complementary evidence comes from the case of an amnesic man who suffered a lesion involving the retrosplenial cortex and cingulum bundle (Valenstein et al. 1987). A subsequent PET study revealed evidence of hypometabolism in the thalamus, while the medial temporal region appeared normal (Heilman et al. 1990). It was concluded that the amnesia in this patient was a consequence of the disconnection of anterior thalamic pathways (Heilman et al. 1990). A similar explanation could account for a patient who suffered a circumscribed haemorrhage in the left retrosplenial area and cingulum bundle (Von Cramon 1992). A mild, but perceptible, loss of verbal learning and memory was observed which disappeared after six months (von Cramon 1992).

Other evidence comes from a person who displayed a persistent impairment in the recall of verbal material, but a sparing of recognition (Hanley et al. 1994). The impairment followed an anterior communicating artery aneurysm which resulted in left hemispheric damage in the anterior thalamus and caudate nucleus (Hanley et al. 1994). In another informative case a woman suffered an infarct centred on the anterior thalamic nuclei and the genu of the right internal capsule. She developed a severe anterograde amnesia for verbal and nonverbal information (Schnider et al. 1996) which was characterised by a failure to use contextual information so that her recognition performance was marred by high levels of false positives. Her performance improved dramatically to within normal limits when tested with nonsense words and nonsense designs, stimuli for which contextual information (i.e. remembering) would be of less value but familiarity might be sufficient (Schnider et al. 1996). While this pattern of performance clearly fits with the current proposals it should be noted that the infarct probably disconnected other potentially important fibre tracts. Finally, the production of bilateral radiofrequency lesions in the anterior nuclei induced a loss of memory in a patient with chronic depression (Mark et al. 1970). It was possible to confirm the placement of the lesions following the suicide of the patient six weeks after the surgery. The lesion destroyed the whole of the anterior nuclear complex "but essentially no cells of nucleus ventralis anterior or nucleus dorsomedianum" (Mark et al. 1970). Although the patient refused to take memory tests she expressed concern about her recent memory loss and she had difficulty in remembering the location of her hospital room. Although the authors regarded the memory loss as transient this could not be verified.

4.3 Electrophysiological Studies:

Neuronal recording studies provide strong support for the idea of a division of function within the medial temporal lobe between the perirhinal and hippocampal cortices. Indeed, electrophysiological evidence of the importance of perirhinal cortex rather than the hippocampus for judging stimulus familiarity preceded lesion evidence (Brown et al. 1987). A number of studies have now examined neuronal responses in anterior inferior temporal cortex, including perirhinal cortex, during the performance by monkeys of recognition tasks using large sets of stimuli (Brown et al. 1987; Eskander et al. 1992; Fahy et al. 1993a; Li et al. 1993; Lueschow et al. 1994; Miller et al. 1993; Riches et al. 1991; Sobotka & Ringo 1993, 1994; see for reviews, Brown 1996; Brown & Xiang 1998; Ringo 1996). The tasks employed have been either variants of DMS or variants of a serial (running) recognition memory task (Gaffan 1974). These studies have confirmed that many neurones in perirhinal cortex respond maximally to first presentations of visual stimuli, but less to subsequent presentations. Hence the re-occurrence of a particular stimulus is signalled by a decrease in the neuronal response to that stimulus. Similar responses have been found in the perirhinal cortex of the rat (Zhu et al. 1995a). It should be noted that in monkeys neurones that increase their responses when a stimulus re-occurs are found even less often than might be expected by chance (Xiang & Brown 1998).

These changed responses with stimulus repetition are sufficient to solve recognition memory tasks such as DMS and DNMS as commonly tested in monkeys. Thus: (i) a single exposure to a stimulus is sufficient to cause a change in responsiveness, i.e. one trial learning; (ii) for many neurones, the response change is found even when a long period of time (for example, 24 hours) has elapsed and/or many (hundreds of) presentations of other stimuli have intervened between the first and the next presentation of a stimulus; and (iii) the effect is highly stimulus-specific, i.e. a neurone that responds weakly to a stimulus that has been seen before still responds strongly to novel stimuli, though there may be a limited amount of stimulus generalisation (Brown 1996; Lueschow et al. 1994; Ringo 1996). Moreover, individual neurones independently signal different types of information of potential use to the judgement of prior occurrence (Fahy et al. 1993a; Zhu et al. 1995a). Thus there are `recency neurones' that encode whether a stimulus has been seen recently irrespective of whether it has been seen many or few times previously. Other, `familiarity neurones' encode the relative familiarity of stimuli (i.e. whether they have been seen many or a few times previously) without regard to whether they have been seen recently. A third category responds best to novel stimuli or first presentations of unfamiliar stimuli that have not been encountered recently (Xiang & Brown 1997, 1998). All three types of neurones are found in anterior and medial temporal cortex (area TE) and in entorhinal cortex as well as in perirhinal cortex (Xiang & Brown 1997, 1998). Neurones have been reported (Li et al. 1993) whose response changes progressively as initially unfamiliar stimuli are successively repeated (and presumably become increasingly familiar). Other types of neuronal activity changes such as incremental responses (Miller & Desimone 1994) and increased firing after stimulus offset (delay activity: Fuster & Jervey 1981) are also seen in these regions, and these may facilitate or provide alternative means of solving specific types of recognition memory tasks (Brown 1996).

Carefully controlled experiments have established that the described response changes are not artefactual: they cannot be explained by alterations in eye or body movements, or motivational or attentional factors (see Brown 1996). Importantly, they are found not only during the performance of recognition memory tasks, but also when stimuli are shown without the animal being required to make a particular behavioural response and using types of stimuli not employed in the animal's training (Brown 1996; Riches et al. 1991). Accordingly, the response changes are an endogenous rather than a training-induced property of the neurones.

Not all visually responsive neurones recorded in perirhinal cortex change their response with stimulus repetition - many respond equivalently even after the same stimulus has been shown many times. The latter type of neurone encodes information concerning the physical characteristics of the stimuli and so may contribute to stimulus identification. Nevertheless, the incidence of neurones whose responses do change is about 20-25% of the whole neuronal population (Brown 1996), a finding that has been exploited in c-fos studies (Section 4.4). Neurones with changing and neurones with unchanging responses are found intermingled, so that comparisons between cells with the two types of response could provide information concerning the previous occurrence of a stimulus. Further, as the responsiveness of both types of neurones varies with the physical characteristics of the stimulus, and the degree of response decrement and the time for which it lasts varies amongst neurones whose responses change, population measures can potentially provide a sensitive measure of the past occurrence of a particular stimulus (both of its recency of occurrence and of its familiarity).

There is also electrophysiological and ablation evidence that the perirhinal region is involved in the learning of visual paired associates (Higuchi & Miyashita 1996; Miyashita et al. 1996; Murray et al. 1993; Okuno & Miyashita 1996). Thus electrophysiological evidence supports a role for perirhinal cortex in recording knowledge concerning individual stimuli and of simple associations between such stimuli. Information concerning the prior occurrence of individual stimuli is also available in the inferior temporal cortex adjacent to perirhinal cortex, but in that region it is likely to be restricted to the visual modality. Perirhinal cortex receives multimodal sensory input (Burwell et al. 1995) and so may serve a similar function for information not confined to a single modality, although present evidence is largely confined to the visual modality. Information about prior occurrence is also available within entorhinal cortex (Fahy et al. 1993b; Suzuki et al. 1995; Xiang & Brown 1998). In contrast, neurones that change their response after a single occurrence of an individual stimulus are uncommon in the hippocampus of monkeys (Brown et al. 1987; Miyashita et al. 1989; Riches et al. 1991; Rolls et al. 1993; Xiang & Brown 1998) and rats (Zhu et al. 1995a) - a finding confirmed in c-fos studies in the rat (Wan et al. 1997; Zhu et al. 1995b, 1996). Thus such perirhinal and entorhinal information is not necessarily passed on to hippocampal neurones in spite of the numerous interconnections between these regions.

Contrasting with the findings for individual stimulus repetition, there is much evidence that the activity of hippocampal neurones reflects features of the spatial environment or spatial arrangements of stimuli (Eichenbaum et al. 1994; Eifuku et al. 1995; Gothard et al. 1996; O'Keefe 1993; O'Keefe & Nadel 1978). In particular, individual neurones increase their firing when a rat is located in a particular part of an environment (`place fields'). Such place fields are influenced by the presence of defining features of that environment, including its size and shape (O'Keefe & Burgess 1996). In the monkey, increases in firing have been recorded when the animal directs its gaze to a particular part of the environment (`view fields') (Rolls & O'Mara 1995). Additionally, changes in hippocampal responses related to a combination of the relative familiarity of a stimulus and the place in which it was shown have been described (Miyashita et al. 1989) - as they have in entorhinal cortex (Suzuki et al. 1995). Although hippocampal neurones also respond to discrete stimuli and during paradigms that are not spatial (e.g. Brown 1982; Eichenbaum et al. 1994; Eifuku et al. 1995; Otto & Eichenbaum 1992; Thompson 1990), electrophysiological recordings have not yet established that the response characteristics of hippocampal neurones could provide satisfactory explanation for the properties of episodic memories. Spatial information is important to episodic memory, but it does not encompass the whole of the information contained in the context and content of all episodic memories. Nevertheless, the anatomical connectivity of the hippocampus is ideal for making the widely distributed associations (Brown 1990; Lorente de No 1934) between the many different contemporaneously experienced stimuli comprising an event. Such associations are necessary to the formation of an episodic memory.

Recordings in the monkey medial thalamus have revealed neurones whose responses signal information concerning the prior occurrence of stimuli. Neurones in both the medial dorsal nucleus and the paraventricular midline nucleus responded more strongly to first than to subsequent presentations of stimuli presented during the performance of a recognition memory task (Fahy et al. 1993b). Such findings are consistent with the suggested functional link between perirhinal cortex and the medial dorsal nucleus of the thalamus, though the numbers of such neurones were small both in absolute and in percentage terms. There are no reported recordings from the anterior thalamic nuclei during the performance of recognition memory tasks. However, neuronal responses in the anterior thalamic nuclei in the rat can code spatial information, particularly about the direction of the head (Mizumori & Williams 1993; Taube 1995). Thus thalamic recordings again support the apportionment of function between the hippocampal - anterior thalamic and the perirhinal - medial dorsal thalamic axes, with the latter underlying the detection of the prior occurrence of a stimulus (its sense of familiarity, `knowing' that something has occurred previously) and the former being involved in the processing of important components, including the contextual information concerning events as is necessary for `remembering' or `recollecting' a prior occurrence.

4.4 Brain Activation Studies: c-fos

The expression of the immediate early gene c-fos provides a potential marker for changes in neuronal activity (Herdegen 1996), and so can be used to identify regions that respond to specific experimental manipulations. We have used this technique to help map brain regions involved in recognition (Zhu et al. 1995b; 1996). In the first of these studies, counts of fos stained nuclei were compared in the brains of rats shown novel objects and in the brains of rats shown highly familiar objects (Zhu et al. 1995b). The objects were placed behind a glass screen, so limiting the task to visual processes. Four of the eight sites examined showed significantly higher levels of fos protein in the brain exposed to novel stimuli. These sites were the occipital cortex, area TE, perirhinal cortex and the anterior cingulate cortex. No differences were observed in the entorhinal cortex, the diagonal band of Broca, the medial dorsal nucleus of the thalamus or the hippocampus (Zhu et al. 1995b). These cortical sites of increased activation correspond closely to those described in a human PET study comparing exposure to familiar and unfamiliar visual stimuli (Vandenberghe et al. 1995). Increased blood flow was found in the lateral anterior temporal neocortex, the medial temporal pole, and the anterior cingulate cortex following exposure to the unfamiliar stimuli (Vandenberghe et al. 1995).

A shortcoming with the c-fos result concerns the need to control for any differences in behaviour induced by the presence of novel objects (e.g. eye movements, alertness). For this reason a second study used a within-subjects design (Zhu et al. 1996). On each trial two objects were shown simultaneously to the rat, so that one eye saw only novel objects while the other eye saw only familiar objects.

Figure 6. Testing arrangement for the c-fos `paired-viewing' procedure in which one hemisphere receives direct visual information from an eye seeing novel stimuli while the other eye views familiar stimuli (Zhu et al. 1996)

This `paired-viewing' procedure makes it possible to compare fos levels in the two hemispheres of the same brain. Novel stimuli activated significantly more neurons (as measured by fos) in the perirhinal cortex, area TE, and the ventral lateral geniculate nucleus of the thalamus. No differences were observed in the hippocampus or any other area sampled. Moreover, the number of stained (activated) neurones in the hippocampus was low.

The failure of novel stimuli to evoke significantly increased c-fos products in the hippocampus might be because neurones in this structure do not show activity-related expression of c-fos. This possiblity was excluded by a third study (Zhu et al. 1997) in which exposure to novel objects took place in either a familiar or a novel context. Relative to the perirhinal cortex, the hippocampus (fields CA1-4) showed a fourfold increase of activity for the novel context compared to the familiar context. This result, which contrasts with that for novel compared to familiar objects in perirhinal cortex, provides a double dissociation between the hippocampus and the perirhinal cortex. As a consequence the result concurs with lesion and electrophysiological studies pointing to a qualitative difference between the contributions of the two regions to recognition and contextual (spatial) memory.

In order to explore this difference further a subsequent study compared c-fos activation when rats were exposed either to novel or familiar visual stimuli or to novel or familiar spatial rearrangements of familiar stimuli (Wan et al. 1998). As in previous studies the former condition led to higher c-fos expression for novel than for familiar stimuli in the perirhinal cortex but not in the hippocampus or postrhinal cortex. In contrast, the novel spatial compared to familiar spatial rearrangements resulted in increased c-fos activation in the postrhinal cortex, and while some regions of the hippocampus showed increased (CA1) activation others (dentate gyrus and subiculum) showed decreased activation (Wan et al. 1998). In view of the supposed homology between the postrhinal cortex and the parahippocampal gyrus (Burwell et al. 1995) these findings provide an insight into how novelty per se need not engage the hippocampus, but novelty arising from the mismatch of certain learnt associations in space may do so. This conclusion matches that observed in lesion studies of novelty detection (Honey et al. 1998) and conditional learning (Gaffan & Harrison 1989). It also accords with a recent fMRI study showing raised activation in the parahippocampal gyrus during the presentation of novel words or scenes, but increased activation in the subiculum during retrieval (Gabrieli et al. 1997).

5. The Consequence of Multiple Process Models of Recognition

Up to now recognition has largely been considered as if it were a single process, but many psychologists argue that it involves at least two distinct processes (Gardiner & Parkin 1990; Jacoby & Dallas 1981; Mandler 1980). One of these permits a recognition judgement to be made on the basis of stimulus familiarity (sometimes regarded as `perceptual fluency'). Using this information a subject will have a feeling of `knowing' that they have experienced the test item but may remember no other associated details. The second process is regarded as elaborative as it involves recollecting (`remembering') the experience of the test item. As a consequence it may be recalled along with associated contextual information (Perfect et al. 1996). Familiarity is regarded as an automatic process while recollection is an effortful retrieval process. An alternative view of recognition is that `know' responses just reflect a weaker recognition process than `remember', and that these are part of the same mechanism.

Evidence for this two process account comes from manipulations, such as changing levels of processing and increasing the interval between sample and test, that differentially affect the two classes of response (Gardiner 1988; Rajoram 1993; Tulving 1985). Most importantly, there are conditions that affect `know' responses but not `remember', and vice versa, so producing a double dissociation between the two classes of response (Gardiner & Java 1990; Rajoram 1993). In spite of these dissociations, it has been argued that a two process model may not be necessary and that it is possible to model these dissociations using a single process model within which different criteria are set for `remember' and `know' responses (Donaldson 1996; Hirshman & Master 1997). It should also be noted that some of this debate relates not to the notion of a distinction between recollective and nonrecollective memory, but more to the way that recognition memory is measured (Donaldson 1996). Furthermore, recent analyses of ROCs (receiver operator characteristics) provide further weight to dual process models (Yonelinas 1994, 1997). This final approach is of especial interest as it has led to a comparison (Yonelinas 1997) of recognition judgements for item information and associative information (learnt pairs). This revealed evidence that item recognition and associative recognition rely differentially on recollection and familiarity. While item recognition principally reflected familiarity but also and recollection, the recognition of associative information relied on recollection (Yonelinas 1997). This distinction is especially relevant as it maps onto the proposed differences between the contributions of the perirhinal cortex to recognition (item familiarity) and the hippocampus to recognition (associative recollection).

It is therefore assumed that the hippocampal - anterior thalamic system supports recollective based recognition (remembering) while cortical (perirhinal) systems support familiarity based recognition (knowing). As a consequence, all amnesics will show a marked fall in recollective based recognition while the loss of familiarity based recognition will vary according to the extent of cortical (extrahippocampal) dysfunction. Thus very selective hippocampal system damage may spare those tests of recognition that can be performed effectively by just using relative familiarity. In contrast, more extensive cortical damage, as found in the majority of cases, will disrupt both processes. The proposal that the perirhinal contribution to this form of memory is distinct from that of the hippocampus (as double dissociations can be found) has further implications. Namely, that recollection could occur without familiarity. There is good evidence that this can occur (Jacoby & Dallas 1981; Jones & Gardiner 1990; Mandler 1980), but it would be expected to be unusual given the interlinked flow of afferent sensory information to the rhinal region and hippocampus (Gaffan & Parker 1996; Mishkin & Murray 1994; Suzuki 1996b). From this proposal it can be seen that the two components of recognition are regarded as distinct processes that are independent of one another (Jacoby 1991; Yonelinas & Jacoby 1995) but not mutually exclusive. This latter point is important as tests based on `remember' and `know' decisions carry with them an assumption that the processes are mutually exclusive (subjects can never respond both `know' and `remember'), and for this reason provide an imprecise measure of the two processes. This is because `know' responses reflect familiarity in the absence of recollection, while `remember' responses will contain some items that are both recollected and familiar (see also Donaldson 1996).

One method for testing these possible dissociations within memory has been to assess whether manipulations that lead to similar levels of recognition in amnesics and controls also result in matched levels of recall. In one study Korsakoff subjects received an extended study time so that their word recognition performance was equated with that of control subjects (Hirst et al. 1986). The same extended period of study did not, however, abolish the recall deficit and it was concluded that recall and recognition are differentially affected by amnesia (Hirst et al. 1986). A second study equated recognition performance in amnesics and controls by testing the controls with longer retention delays (Hirst et al. 1988). At these same delays the recall performance of the amnesics was still significantly poorer than that of the controls (Hirst et al. 1988). Similar studies have since, however, been repeated (Haist et al. 1992) and these failed to show a difference between recall and recognition. This led to the conclusion that recall and recognition are tightly linked functions of declarative memory (Haist et al. 1992). The difficulty with these experiments is that they treat recognition as a unitary process, and assume that equating groups across different retention delays is valid. One problem is that `know' and `remember' responses have differential rates of forgetting (Tulving 1985), so limiting such comparisons.

A second study (Knowlton & Squire 1995) used the know/remember distinction to examine recognition performance in a group of 13 amnesics. The amnesics showed a significant loss of both `remember' and `know' responses, but revealed a disproportionate loss of remember responses when compared with control subjects tested with the same retention delay. Although this pattern is consistent with the present proposals, the same study did include some amnesics with relatively selective pathology. The study reported that these cases did not show a different profile of impairment (Knowlton & Squire 1995) as might be predicted (but note the debate over anoxia and covert pathology, Section 3.1). It is, however, the case that the data for these individual subjects were not provided, and shortcomings of the `know'/'remember' distinction have already been discussed.

Single case information comes from an atypical Wernicke-Korsakoff subject who showed remarkable sparing on tests of recognition (Parkin et al. 1993), suggestive of a selective loss of recollective based recognition. She was tested using the know/remember subjective distinction, and although she produced the predicted increase in `know' responses she only showed a very modest drop in `remember' responses. Other evidence comes from an amnesic whose performance on different tests of recognition, following anterior communicating artery aneurysm, was best described by the extent to which familiarity information could be utilised (Parkin et al. 1994). Although this case supports the current proposals it was not possible to determine the locus or extent of the brain injury. Other support comes from studies showing that the recognition performance of Korsakoff amnesics depends on judgements of trace strength i.e. familiarity (Huppert & Piercy 1978).

Taken together, these studies provide general support for the current proposals but they fail to reveal the pattern of recognition loss following selective hippocampal system damage. The finding that amnesics with selective pathology can perform the Warrington Recognition Memory Test (RMT) within normal limits (Aggleton & Shaw, 1996) only accords with the present views if it is assumed that the RMT can be performed accurately using just familiarity information. Some support for this assumption comes from the discovery that performance of the Faces subtest of the RMT (the Words subtest was not examined) is not influenced by extreme switches in context between the sample phase and the test phase (Parkinson & Aggleton 1994) even though the same switches markedly impair recall (Martin & Aggleton 1993; Godden & Baddeley 1975) and, hence, should disrupt recollective processes.

At this point it is necessary to consider the relationship between familiarity and priming, particularly repetition priming. As familiarity (a feeling of 'knowing') is an essentially explicit, conscious form of memory, whereas priming is fundamentally implicit, these two processes cannot have identical substrates, i.e. 'know' responses cannot be a direct consequence of priming mechanisms and nothing else. Nevertheless, many descriptions of 'knowing' regard it as reflecting increased perceptual fluency, i.e. priming. Moreover, in the present model familiarity as well as priming is regarded as an automatic consequence of passive exposure to stimuli. Accordingly, processes leading to priming and to familiarity will normally co-occur, yet the two can be dissociated. Thus priming is often preserved in amnesics even though they are typically poor at tests of recognition (Cermak et al. 1985; Reed et al. 1997; Schacter et al. 1991; Warrington & Weiskrantz 1974), including the RMT. This distinction is highlighted in an amnesic E.P. who showed chance performance on a range of recognition tasks i.e. could not use familiarity information, yet displayed intact priming on matched tasks (Reed et al. 1997). Other evidence comes from the loss of `know' responses in many amnesics (Knowlton & Squire 1995), which contrasts with the preservation of other forms of implicit memory.

These findings leave open the possibility that priming and familiarity share the same initial stages of processing, but that familiarity requires additional, separate mechanisms that give rise to subjective appreciation ('knowing') and allow explicit guidance of choice behaviour. Two possibilities concerning these additional mechanisms are suggested by the data available. One is that processing involving the integrity of perirhinal cortex is necessary for familiarity but not for priming: this is consistent with the studies reviewed above, including those of patient EP (Reed et al. 1997). The other requires that the absence of contextual information (remembering) renders very difficult the use of primed information to aid recognition, even though priming is intact as measured by other means. However, this second suggestion leaves unanswered the problem of why this implicit information cannot be utilised even in a forced-choice paradigm (Reed et al. 1997). Similarly, it does not explain why monkeys with hippocampal system damage can still perform at normal or near normal levels on DNMS tasks, and how some amnesics can perform within normal limits on certain tests of recognition (as both have lost contextual information). Thus it seems necessary to adopt the former of these suggestions and conclude that, unlike repetition priming, feelings of familiarity depend upon processing involving perirhinal cortex.

This conclusion may well be relevant to a study using the `process dissociation procedure' (Jacoby 1991) which is designed to separate recollective (explicit) processes from implicit aspects of recognition. It was found that amnesics showed a disproportionate loss in the use of recollection as a basis for recognition (Verfaellie & Treadwell 1993). This component of the study supports the current proposals as such a loss would reflect the core loss of episodic memory. The amnesic subjects were, however, as likely as control subjects to use perceptual fluency (Verfaellie & Treadwell 1993). This latter finding is also consistent with the current proposals if the measurement of `familiarity' in the process dissociation procedure corresponds to implicit memory (which may be spared). If, on the other hand, this measure reflects perirhinal familiarity then the results are at odds with the current proposals, unless the pathologies were unusually selective. It should also be added that the findings of Verfaellie & Treadwell (1993) have been challenged (Roediger & McDermott 1993; but see Verfaellie 1994), as has the validity of the process dissociation procedure (Dehn & Engelkamp 1997; Ratcliff et al. 1995).

A final consequence of the multiple process model of recognition is that it can unify different descriptions of the deficit that follows selective hippocampal - anterior thalamic damage. The current proposals argue that such damage will spare familiarity based recognition, but impair recollection based processes in recognition and recall. This will result in a loss of associative memory (Yonelinas 1997) and, hence, source memory. This prediction is supported by a study confirming that contextual knowledge for items in a recognition test is consistently higher for `remember' items than for `know' items (Perfect et al. 1996). Thus the current proposals are consistent with those of Gaffan (1991; 1992a; 1994b) i.e. that the hippocampus - anterior thalamic axis is required for the creation of episodic (associative) scenes that can heighten discriminability and so aid retrieval of the to-be-remembered item.

6. Diencephalic amnesia and recognition memory

Amnesia associated with diencephalic damage typically results in a severe loss of recognition. It is therefore assumed that there is a loss of both familiarity and recollective based recognition. While the recollection deficit arises from dysfunctions in the hippocampal - anterior thalamic system, the source of the familiarity deficit must be elsewhere. Experimental evidence from behavioural studies of both monkeys (Aggleton & Mishkin 1983b; Parker et al. 1997; Zola-Morgan & Squire 1985a) and rats (Hunt & Aggleton 1991; Mumby et al. 1993) points to the importance of the medial dorsal nucleus of the thalamus. Lesions in this region impair DNMS acquisition and performance, which heavily taxes familiarity judgements. Cells in the medial (magnocellular) portion of the medial dorsal nucleus and in the adjacent midline nuclei show decremental responses to familiar visual stimuli (Fahy et al. 1993; see section 4.3). Moreover, the medial dorsal nucleus is very often affected in diencephalic amnesic syndromes as gliosis is consistently observed in Korsakoff's disease (Victor et al. 1971), while paramedial thalamic infarcts disconnect both frontal and temporal interactions (Graf-Radford et al. 1990; Malamut et al. 1992). While there are two cases of lesions largely confined within the medial dorsal nucleus that had no apparent effect on memory (Kritchevsky et al. 1987), these only involved about 15% of the nucleus and so do not provide a conclusive test of the present proposals.

The medial dorsal thalamic nucleus also receives inputs from the perirhinal cortex. Direct projections run to the magnocellular portion of the medial dorsal nucleus via the inferior thalamic peduncle (Aggleton et al. 1986a; Russchen et al. 1987). In the light of the current proposals it is noteworthy that these connections do not use the fornix, nor does the perirhinal cortex appear to project to the anterior nuclei of the thalamus (Aggleton & Saunders 1997). The magnocellular portion of the medial dorsal nucleus has dense projections to the medial and orbital frontal cortices, and removal of these regions also impairs DNMS performance in monkeys (Bachevalier & Mishkin 1986). In contrast, removal of the dorsolateral prefrontal cortex, which receives its inputs from the lateral (parvocellular) portion of medialis dorsalis, does not affect DNMS performance (Bachevalier & Mishkin 1986). Furthermore lesions of the uncinate fascicle, which connects the temporal association cortex with the prefrontal cortex, have no effect on DMS (Gaffan & Eacott 1995). These findings all suggest that one set of key perirhinal outputs for recognition are those to the thalamic nucleus medialis dorsalis (Parker et al. 1997) and thence to the prefrontal cortices. These perirhinal outputs are unlikely, however, to be the sole route as the magnitude of the medialis dorsalis recognition deficit appears less than that associated with perirhinal damage (Aggleton & Mishkin 1983b; Parker et al. 1997). This indicates that other perirhinal outputs contribute to recognition, but that none of these other routes are individually critical. Candidates include the projections to prefrontal cortex, entorhinal cortex (Meunier et al. 1993) and those back to the inferior temporal cortex (in the case of visual recognition).

In view of the fact that the major ouput from medialis dorsalis is to the prefrontal cortex it is relevant that frontal damage in humans can sometimes disrupt recognition, but that this is not an invariant finding (Aggleton & Shaw 1996; Schacter 1987; Shimamura et al. 1990). This is highlighted by a recent survey (Wheeler et al. 1995) which concluded that frontal lobe damage could impair recognition although the effects are smaller than those for free recall. It should, however, be noted that dorsolateral prefrontal lesions in monkeys do not disrupt DNMS performance, and it is this region that is involved in many of the human cases. Clinical cases with extensive medial and orbital damage are rarer, and it is possible that such damage might be sufficient to impair recognition consistently. The relative mildness of the frontal recognition impairment may also reflect the presence of projections from the medial dorsal nucleus back to the rostral perirhinal cortex (Markowitsch et al 1985) that might aid recognition. The mild frontal deficit in the judgement of previous occurrence can be contrasted with the more robust failure to discriminate relative recency or source information (Schacter 1987; Shimamura et al. 1990). Evidence that recency judgements and recognition judgements are not based on the same information has been reported elsewhere (Brown 1996; Shaw & Aggleton 1995) and, taken together, these findings suggest that while the prefrontal cortex may receive information concerning familiarity (either directly from the temporal lobes or from the thalamus) this need not be critical for recognition. Indeed, recent PET studies provides evidence of a dissociation between temporal lobe and dorsolateral prefrontal activity, with temporal lobe activity being highest for novel stimuli and prefrontal activity being highest for novel rearrangements of familiar stimuli (Dolan & Fletcher 1997; see also Tulving et al. 1996).

The prefrontal cortex (like the entorhinal cortex) is, in fact, in a privileged position as it receives both familiarity information and source information. This source information is presumably associated with the direct inputs to the prefrontal cortex from the hippocampus, along with inputs from the medial portions of the thalamus, including the anterior thalamic and midline nuclei (Kievit & Kuypers 1977). Brain activation studies also indicate that the frontal lobes have a specific involvement in recollective (Dolan & Fletcher 1997; Wilding & Rugg 1996) and retrieval (Rugg et al. 1996) aspects of recognition. The key difference is that frontal damage does not reproduce the temporal lobe or diencephalic amnesic syndromes so that its involvement is often not critical. To account for this it is assumed that the hippocampal - anterior thalamic axis has both a diffuse frontal extension as well as an important reciprocal component from the thalamus back to the temporal lobe and hippocampus. Activity in this latter system is sufficient to ameliorate some of the effects of prefrontal damage on standard tests of recall and associative recognition. At the same time, the prefrontal cortex enables the most effective recall strategies to be employed.

7. Subdivisions of amnesia

The present proposals run contrary to the traditional division between temporal lobe and diencephalic amnesia, and predict that the core symptoms should appear very similar. Indeed, when the problems of matching the severity of the amnesia among different cases and allowing for differential damage to additional structures (e.g. frontal cortex) are taken into consideration, the evidence for qualitative differences in the core features of these amnesias seems weak (Weiskrantz 1985; Zola-Morgan & Squire 1993). Initial evidence of differences in forgetting rates between temporal lobe and diencephalic amnesia (Huppert & Piercy 1979; Squire 1981) failed to survive later scrutiny (Freed et al. 1987; Freed & Corkin 1988; McKee & Squire 1992). Although some diencephalic amnesics do show a greater loss of short term memory and contextual cues (Parkin 1984; Parkin et al. 1990), these differences often relate to Korsakoff's syndrome which results in frontal dysfunction. This may also account for the loss of short term memory (Leng & Parkin 1989; Cave & Squire 1992). Evidence of a greater failure to use temporal context information is more difficult to resolve (Parkin & Hunkin 1993), but it may well prove to be a combined effect of frontal dysfunction (Shimamura et al. 1990) and a loss of recollective information concerning the learning episode.

Up to now we have sought to emphasize the distinction between a hippocampal - fornix - anterior thalamic system and a perirhinal - medial dorsal thalamic system. It is, of course, the case that the perirhinal cortex is a major afferent source to the entorhinal cortex and thence to the hippocampus. In view of evidence showing that the perirhinal cortex is important for knowledge concerning objects (Suzuki 1996), principally their familiarity and whether they have been associated with other discrete visual inputs, it can be assumed that this route normally provides object (item) related information to the hippocampus that may be retained in episodic memory, although other routes can be used following brain pathology (Section 3.3). The normal process then involves setting the to-be-remembered item (or items) within its episode or context, and for this association to be possible the hippocampus needs to receive spatial/contextual information. In the primate brain the most plausible route is via the parahippocampal cortex (Habib & Sirugu 1987; Maguirre et al. 1996; Suzuki & Amaral 1994) which permits item - place representations to be formed. The situation in the rat brain may be different as the postrhinal cortex does not appear to be necessary for some spatial tasks (Aggleton et al. 1997). If, as is assumed, the anterior thalamic nuclei are vital for episodic memory then these nuclei must interact with those classes of information disrupted by amnesia, including memory for discrete items. Thus, it is to be expected that the pathway connecting the perirhinal cortex - entorhinal cortex - hippocampus - fornix - anterior thalamic nuclei forms the route by which discrete item information is made available for recall.

The assumption that the hippocampus receives object related information from the perirhinal cortex fits with the notion that damage to the hippocampal - anterior thalamic system can disrupt recognition when familiarity information is not available. Further evidence comes from the additive effects of anterior thalamic damage upon posterior medial thalamic damage to DNMS performance by monkeys (Aggleton & Mishkin 1983a, 1983b). Similarly, fornix lesions on their own have very little effect on the standard DNMS task, but they are able to exacerbate the effects of cutting the ventral amygalofugal pathway/temporal stem which disconnects rhinal projections to the medial thalamus (Bachevalier et al. 1985b). Similar evidence has come from studies of rats showing that fornix damage can accentuate the recognition deficit associated with perirhinal damage (Ennaceur & Aggleton 1998; Wiig & Bilkey 1995). In all of these instances, the fornix/anterior thalamic lesion will have involved efferents from the entorhinal cortex as well as the subiculum. The loss of this entorhinal information may contribute to these additive effects.

8. Final Comments:

This target article has been concerned with extending the functional hippocampal system and showing how this relates to the pathology underlying amnesia. What has not been discussed is why this additional diencephalic processing is required. Lesion studies have been relatively unhelpful in this regard as damage to the anterior thalamic nuclei seems to mimic the effects of hippocampectomy so closely . Other approaches are required, and of these electrophysiological studies may prove to be especially valuable. An example of this is the work of Gabriel and his colleagues (Gabriel 1993). Using a discriminative avoidance task in which one tone predicts an avoidable shock (S+) and another tone does not predict shock (S-), training-induced changes in neural activity have been found in both the anterior thalamic nuclei and the posterior cingulate cortex (Gabriel 1993). These consist of increased firing to the conditioned stimuli and greater discriminative firing to S+ than to S-. As both changes are found in the anterior thalamic nuclei and the upper layers of the posterior cingulate cortex (Gabriel 1993), regions which are closely interconnected, it has been proposed that they reflect an interlinked system. Because this training-induced neuronal plasticity is not observed until late in training, when the association is well learnt, it has been described as a `primacy' system which holds primary or original encodings even after more recent information has been obtained. As might be predicted, lesions of the anterior thalamus - posterior cingulate cortex leave initial acquisition of the avoidance task intact but affect final levels of performance (Gabriel 1993). In contrast, the adjacent medial dorsal nucleus and the interconnected anterior cingulate cortex show discriminative activity from the first session in which behavioural discrimination is observed, and are seen as more important in the initial learning stages (a `recency ` system).

While this model system of avoidance learning offers a means of assessing how the anterior thalamic nuclei can interact with cortical regions in a mnemonic capacity, the basic avoidance task is unlikely to provide a direct measure of episodic memory and hence is of limited application. Consistent with this is the finding that rabbits with hippocampal lesions can show normal acquisition and performance of the avoidance task (Gabriel 1993), and hippocampal lesions do not disrupt training induced plasticity in the anterior thalamic nuclei. Conversely, the cingulate cortices are seen as vital for the avoidance task yet are not critical components of the episodic memory system outlined in this review. There is evidence, however, that the hippocampus modulates overall levels of activity in the anterior thalamic nuclei during the avoidance task. Furthermore, these interactions may be influenced by context and this may depend on hippocampal activity. Evidence comes from the finding that hippocampal lesions in rabbits attenuate the effects of a context shift on extinction of the conditioned avoidance response (Gabriel 1993). It can be seen that such preparations are moving closer to the demands of an episodic memory system and so may help to address more directly the important issue of why diencephalic relays are required in normal episodic memory.

An underlying assumption in this review is that the critical contribution of the anterior thalamic nuclei to episodic memory will involve not only its hippocampal and mamillary body inputs but also its other afferents. A similar assumption applies to the mamillary bodies, that is these structures are contributing something new and not merely passively processing hippocampal outputs. In the case of the anterior thalamic nuclei one potentially important input is the ascending cholinergic projection from the tegmentum. Preliminary evidence comes from the finding that acquiring the discriminative avoidance learning task (Gabriel 1993) leads to increased muscarinic acetyl choline binding in the anterior thalamic nuclei, and this correlates with the appearance of training induced changes in neural activity (Vogt et al. 1991). A future task will be to look systematically at these and other inputs to the anterior thalamic nuclei and mamillary bodies, and so better determine the nature of the contribution of these structures to the hippocampal - anterior thalamic axis. As a consequence, the analysis of lesions will continue to refine our understanding of the critical pathologies underlying aspects of amnesia, but quite different techniques will be required to identify how these different structures contribute to the encoding and maintenance of episodic memory.

Acknowledgements

The research of both the authors has been supported by the Medical Research Council (U.K.) and the Wellcome Trust. The authors wish to thank W. Macken and the helpful comments of the referees.

References

Aggleton, J. P., Desimone, R. & Mishkin, M. (1986a) The origin, course, and termination of the hippocampo-thalamic projections in the macaque. Journal of Comparative Neurology 243:409-421.

Aggleton, J. P., Hunt, P. R., Nagle, S. & Neave, N. (1996) The effects of selective lesions within the anterior thalamic nuclei on spatial memory in the rat. Behavioural Brain Research. 81: 189-198.

Aggleton, J. P., Hunt, P.R. & Rawlins, J. N. P. (1986b) The effects of hippocampal lesions upon spatial and non-spatial tests of working memory. Behavioural Brain Research 19:133-146.

Aggleton, J. P., Hunt, P. R. & Shaw, C. (1990) The effects of mammillary body and combined amygdalar-fornix lesions on tests of delayed non-matching-to-sample in the rat. Behavioural Brain Research 40:145-157.

Aggleton, J.P., Keen, S., Warburton, E. C. & Bussey, T.J. (1997) Extensive cytotoxic lesions of the rhinal cortices impair recognition but spare spatial alternation in the rat. Brain Research Bulletin 43:279-287.

Aggleton, J. P., Keith, A.