HIPPOCAMPUS 20:1315–1326 (2010)

Predictive, Interactive Multiple Memory Systems

Richard N. Henson* and Pierre Gagnepain

ABSTRACT: Most lesion studies in animals, and neuropsychological in the pattern of brain activity during various memory
and functional neuroimaging studies in humans, have focused on finding tasks, as measured for example by functional magnetic
dissociations between the functions of different brain regions, for exam-
ple in relation to different types of memory. While some of these disso- resonance imaging (fMRI), are often cited as further
ciations can be questioned, particularly in the case of neuroimaging evidence in support of multiple memory processes.
data, we start by assuming a ‘‘modal model’’ in which at least three dif- These fMRI claims most often relate to the medial tem-
ferent memory systems are distinguished: an episodic system (which poral lobes (MTL), particularly dissociations between
stores associations between items and spatial/temporal contexts, and the hippocampus and surrounding perirhinal cortex and
which is supported primarily by the hippocampus); a semantic system
(which extracts combinations of perceptual features that define items, parahippocampal gyri (e.g., Diana et al., 2007). Here
and which is supported primarily by anterior temporal cortex); and we evaluate these dissociations critically, at least within
modality-specific perceptual systems (which represent the sensory the domain of fMRI studies of human recognition
features extracted from a stimulus, and which are supported by higher memory (see Brown et al., 2010; Montaldi and Mayes,
sensory cortices). In most situations however, behavior is determined by this issue, for evidence from animal and human lesion
interactions between these systems. These interactions reflect the flow of
information in both ‘‘forward’’ and ‘‘backward’’ directions between studies), before proposing a new framework that
memory systems, where backward connections transmit predictions about emphasizes instead associations rather than dissociations
the current item/features based on the current context/item. Importantly, between the activity of different brain regions.
it is the resulting ‘‘prediction error’’—the difference between these We begin by briefly reviewing cognitive and neural
predictions and the forward transmission of sensory evidence—that drives theories of recognition memory, before illustrating
memory encoding and retrieval. We describe how this ‘‘predictive inter-
active multiple memory systems’’ (PIMMS) framework can be applied to some criteria for establishing neuroimaging support for
human neuroimaging data acquired during encoding or retrieval phases of multiple memory systems, based on patterns of fMRI
the recognition memory paradigm. Our novel emphasis is thus on associa- response in hippocampus and perirhinal cortex associ-
tions rather than dissociations between activity measured in key brain ated with recollection and familiarity. Even if no study
regions; in particular, we propose that measuring the functional coupling has yet met these criteria fully, we nonetheless assume a
between brain regions will help understand how these memory systems
interact to guide behavior. V 2010 Wiley-Liss, Inc.
C
modal, multiple-memory-systems model in which rec-
ollection and familiarity depend on the ‘‘forward’’ and
KEY WORDS: recollection; familiarity; priming; memory systems ‘‘backward’’ coupling between systems. We call this
model PIMMS, which stands for ‘‘predictive interactive
multiple-memory systems,’’ where the novel ‘‘pre-
dictive’’ and ‘‘interactive’’ aspects refer to the idea that
INTRODUCTION
‘‘higher’’ systems are constantly trying to predict activity
in ‘‘lower’’ systems on the basis of past experience. We
The debate over whether recollection and familiarity are qualitatively sketch how PIMMS might explain the common pat-
different memory processes reflects a broader debate over the fractiona- terns of recollection- vs. familiarity-related activity in
tion of memory, such as, whether there exist distinct memory systems hippocampus and perirhinal cortex, before generalizing
(e.g., Schacter and Tulving, 1994). While multiple memory processes/ to more everyday situations, and making predictions for
systems may be helpful at one level of description—for example, to future experiments. While it becomes more difficult to
explain behavioral dissociations between different types of memory test such fully-interactive models based on behavioral
test—they may not necessarily have spatially-segregated or independ- data or even local patterns of fMRI responses, we sug-
ently-operating implementation in the brain. Nonetheless, dissociations gest that they can also be tested by examining changes
in the functional coupling between brain regions.
MRC Cognition and Brain Sciences Unit, Cambridge, England, United
Kingdom
Additional Supporting Information may be found in the online version of Theories of Recognition Memory
this article.
Grant sponsor: UK Medical Research Council; Grant number: WBSE People’s ability to recognize a stimulus that they
U.1055.05.012.00001.01. encountered previously has been hypothesized to
*Correspondence to: Dr. Richard N. Henson, MRC Cognition and Brain entail two qualitatively different types of memory:
Sciences Unit, 15 Chaucer Road, Cambridge, CB2 7EF, UK. recollection and familiarity (e.g., Mandler, 1980). In
E-mail: [email protected] keeping with other articles in this special issue, we use
Accepted for publication 17 June 2010
DOI 10.1002/hipo.20857 these terms to refer to psychological processes that
Published online 6 October 2010 in Wiley Online Library both give rise to conscious memory for past exposure
(wileyonlinelibrary.com). to a stimulus, differing in whether (recollection) or

C 2010
V WILEY-LISS, INC.
1316 HENSON AND GAGNEPAIN

not (familiarity) that memory includes details of the episodic tion between the same set of brain regions. Conversely though,
context surrounding that exposure. Experimental assay of recol- spatially-distinct neural systems (for the same type of memo-
lection and familiarity includes objective tests of aspects of that randa) would seem to imply different memory processes. Hence,
prior context (or ‘‘source memory’’; Johnson et al., 1993) or evidence from neuroimaging (or other brain data) that different
subjective judgments of context retrieval (such as ‘‘remember’’ brain regions are differentially involved in experimental condi-
vs. ‘‘know’’ judgments; Tulving, 1985). Rather than discuss the tions presumed to differ only in terms of recollection and famili-
pros and cons of such experimental methods (see instead Mon- arity would support dual-process models of recognition memory.
taldi and Mayes, this issue; Wixted et al., 2010), we assume Several researchers have mapped multiple memory systems
here that recollection has occurred when there is evidence of re- onto the brain, focusing on the MTL. Some are based on
trieval of at least one type of extrinsic context (e.g., spatial or specific computational principles (e.g., Norman and O’Reilly,
temporal information, or internally-generated thoughts during 2003); others are verbal accounts (e.g., Aggleton and Brown,
prior exposure). Note that we do not include memory for con- 1999). A common theme is that recollection arises from a con-
text intrinsic to the stimulus, such as perceptual details not textual/relational/pattern-separating system in the hippocampus,
repeated at test (e.g., Staresina and Davachi, 2008; cf. Wais whereas familiarity arises from item-based/pattern-completing
et al., 2008), which we hypothesize are bound within a seman- cortical systems, such as the perirhinal cortex (see, e.g.,
tic memory system (see later). Norman, 2010; though see also Squire et al., 2007). Before
When fitting behavioral data from recognition memory proposing our own modified multiple memory system view,
paradigms, assumptions are needed about how recollection and we pause to consider the human neuroimaging evidence for
familiarity cooccur or even interact. In one popular model multiple memory systems.
(see Yonelinas et al., 2010), recollection and familiarity are
assumed to be independent. In such ‘‘dual-process’’ models of
Neuroimaging Evidence for Multiple
recognition memory, familiarity is normally assumed to be a
Memory Systems
continuous signal, on which people place response criteria to
make their recognition decision, consistent with signal detection To infer from neuroimaging data that more than one process
theory (SDT, Green snd Swets, 1966). Recollection on the other is engaged in a memory task, Henson (2005) argued that one
hand can be modeled as a discrete probabilistic event (a ‘‘high- must show a ‘‘qualitative’’ difference in the activity of two more
threshold’’ model) or as a second (orthogonal) continuous signal brain regions across three or more experimental conditions.
(in two-dimensional SDT theories, e.g., Rotello et al., 2004). The general idea of a qualitative difference is illustrated in
Others however have argued that the behavioral data are suffi- Figure 1. Assume that one measures the mean event-related
ciently modeled by one-dimensional SDT (see e.g., Wixted fMRI response to three types of trial: (1) correct recognition
et al., 2010). Some of these researchers take this sufficiency to of a studied item together with correct retrieval of its study
indicate that there is only one type of memory process underly- context (C1); (2) correct recognition of a studied item but
ing recognition memory (‘‘single-process’’ theories, such as some failure to retrieve its study context (C2); and (3) correct
‘‘global matching’’ models, Humphreys et al., 1989), such that rejections (CR) of unstudied items. Assume further that one
recollection and familiarity are regarded as quantitatively rather obtains these data from two, independently-defined brain
than qualitatively different. It is important to note however that regions: perirhinal cortex and hippocampus.
one-dimensional SDT models do not necessarily reject the idea The null hypothesis here is that the three conditions differ
of qualitatively different memory processes; they assume simply only along a single dimension of ‘‘memory strength,’’ with a
that, at least for a basic old/new decision, all information relative ordering of activity CR < C2 < C1 (i.e., a single-
retrieved from memory is combined into a single ‘‘strength of process model). An alternative hypothesis is that the conditions
evidence’’ in order for people to make a decision. Nonetheless, differ in terms of recollection (which is greater in C1 than
one take on this ongoing debate between single- and dual-pro- C2) and familiarity (which is greater for C2 than CR). How-
cess models is that behavioral data from the recognition memory ever, the precise mapping between these psychological processes
paradigm are insufficient to resolve this debate (particularly if (memory strength/recollection/familiarity) and neural activity
SDT is generalized to allow unequal signal and noise variances in each brain region (let alone the mapping between neural
or even non-Gaussian distributions; see, e.g., Shimamura, 2010), activity and the fMRI signal) is unknown; it is unlikely to be
and one should instead seek convergent evidence from other linear, for example. Therefore, in order to make any type of
memory paradigms and other types of data, such as those from inference from fMRI data, we make the minimal assumption
neuroimaging (Henson, 2006). that this mapping is monotonic (Henson, 2005).
Here we distinguish between memory processes and memory The pattern of data at the top of Figure 1A is not sufficient
systems, where we take the latter to entail spatially-distinct to reject the null hypothesis, i.e., insufficient to infer separate
neural implementations in the brain. In this sense, dual process processes of recollection (in the hippocampus) and familiarity
models of recognition memory do not imply multiple memory (in the perirhinal cortex). This is simply because the two brain
systems. Recollection and familiarity, for example, could arise regions may have different mappings between memory process
from different computations within the same brain region (e.g., and fMRI signal; though both monotonic, the mapping may
Greve et al., 2010), or, as later, from different types of interac- ‘‘saturate’’ at lower levels of memory strength in the perirhinal

Hippocampus
 PREDICTIVE, INTERACTIVE MULTIPLE MEMORY SYSTEMS 1317

FIGURE 1. Hypothetical patterns of fMRI signal (top row) in evidence of contextual retrieval (C1), analogous to recollection.
perirhinal cortex and hippocampus across three types of trial Assuming a monotonic relationship between fMRI signal and
during the test phase of a recognition memory experiment: correct degree of memory strength (bottom row), only the patterns in
rejections (CR) of unstudied items; correct recognition of studied Panels C and D are sufficient to challenge single-system/process
items without evidence of any contextual retrieval (C2), analogous models of recognition memory (see text for explanation).
to familiarity; and correct recognition of studied items with

cortex than in the hippocampus (bottom of Fig. 1A; see Hen- ordered such that C2 < C1 < CR, whereby the pattern in
son, 2006, for further elaboration, and Squire et al., 2007, for Figure 1C reduces back to an inconclusive pattern like that
a similar argument). It is also possible that increases in memory in Figure 1A. In the more general case, a pattern like that in
strength can produce decreases in the fMRI signal; for example, Figure 1D is necessary to rule out any single-process model.
if the neural basis of familiarity were reduced responsiveness This case corresponds to a ‘‘reversed association’’ (Dunn and
(e.g., Brown and Xiang, 1998; Brown et al., 2010). In Kirsner, 1988), where there is a negative-relationship between
this case, even a cross-over interaction between the three condi- two conditions (C2 and C1), concurrent with a positive-
tions and two regions (Fig. 1B) would not reject the null relationship across two other conditions (C2/C1 vs. CR).
hypothesis. Indeed, the hippocampus and perirhinal cortex may This pattern requires the data from at least two brain regions,
be functionally-coupled in a reciprocal fashion, whereby because then there is no way to reorder the three conditions
increased neural activity in one inhibits activity in the other. consistently across the two regions and maintain monotonic
But what about the pattern in Figure 1C? This pattern psychological-fMRI mappings within each region.
cannot be explained by a monotonic relationship between In fact, we are not aware of any published fMRI study that
memory strength and fMRI signal. Indeed, one common has reported a pattern like that in Figure 1C (or 1D), when
explanation for this pattern is that the hippocampus is involved the regions have been defined independently (i.e., by anatomi-
both in retrieving context, explaining why C1 > C2, and in cal criteria, separate functional data, or orthogonal contrasts on
encoding new items/contexts (even during a recognition test), the same functional data). The studies that have reported
explaining why CR > C2 (e.g., Buckner et al., 2001; Stark similar patterns at retrieval, including our own, have tended to
and Okado, 2003). Another possibility is that this pattern perform post hoc tests across regions that were defined after
reflects a spatial averaging over distinct subfields within the searching for significant correlated effects, which raises the
hippocampus that are differentially involved in encoding and question of statistical bias (Henson, 2006).
retrieval, but beyond than the typical resolution of fMRI (Carr
et al., 2010). Later, we propose an alternative explanation in
A New Multiple Memory Systems Model:
terms of contextual-competition in our PIMMS framework.
From SPI to PIMMS
One might wonder why the data from the perirhinal cortex
are relevant to the above argument; with the nonmonotonic Even if we are not aware of any neuroimaging study to date
pattern in the hippocampus being sufficient to reject the null that has found a qualitative difference in MTL activity associ-
hypothesis of a single dimension of memory strength. This is ated with recollection and familiarity, according to the above
only true if the null hypothesis specifies the relative ordering of criteria, we propose on the basis of other considerations (e.g.,
the C1, C2, and CR conditions; otherwise, one could still neuropsychological, evolutionary, comparative, and computa-
assume a single dimension, but where the conditions were tional, Schacter and Tulving, 1994) that there are multiple

Hippocampus
1318 HENSON AND GAGNEPAIN

FIGURE 2. Tulving’s ‘‘serial-parallel-independent’’ (SPI) model (PIMMS) proposed here (Panel B), where encoding, storage and
of encoding, storage and retrieval of memories (and its extension retrieval are better described as ‘‘interactive-parallel-interactive,’’
in the ‘‘multiple input model,’’ MIM), and its relationship to and where recollection and familiarity entail interactions between
recollection, familiarity and priming (Panel A), in contrast to the multiple memory systems.
new predictive interactive multiple memory system framework

memory systems in the human brain. However, given the high- 2007). This ‘‘multiple inputs model’’ (MIM) has since been
level of anatomical connectivity between brain regions, we extended to include other systems (working and procedural
propose that there is a high-degree of interaction between memory) and to provide a more comprehensive account
memory systems in most situations, which can make their of human memory (e.g., of autobiographical memory); see
dissociation and identification difficult. Eustache and Desgranges (2008).
Perhaps the ‘‘modal’’ multiple memory system model is the se- Given our focus on recollection and familiarity in tests of
rial-parallel-independent (SPI) model developed by Tulving and recognition memory for unrelated, single items, we will focus
Gazzaniga (1995)—Figure 2A. This model assumes at least three on the three-system model in Figure 2B. We distinguish the
memory systems: an episodic system, a semantic system, and systems primarily on their representational content and compu-
one or more perceptual systems. The name SPI refers to the rela- tational principles.i We assume that the main purpose of the
tionship between these systems during encoding, storage and re-
trieval: viz., that encoding is a serial process in which a stimulus
i
is processed first by a perceptual system, then the semantic sys- Since we later propose that these systems operate according to
similar principles, differing only in their representational content
tem and finally the episodic system; storage occurs in parallel, in
(and possibly the type of representation, e.g., sparse or distrib-
that this processing may leave separate memory traces in each uted), one could question whether these are really ‘‘memory’’
system; and retrieval is independent, in that such memory traces systems. We accept that they are not different systems in the
can be accessed independently (depending on the retrieval test, sense of operating according to separate memory principles; or
retrieval cues, etc). According to SPI, recollection is the process even memory systems in the sense of being specialized for
memory (since we believe plasticity is a general property of
of retrieval from the episodic system, while familiarity is the pro-
brain regions, whose specialization depends instead on the na-
cess of retrieval from the semantic system (and the process of re- ture of the information that they code; see also Ranganath,
trieval from a perceptual system is priming). 2010; Cowell et al., 2010). Instead, we use the phrase ‘‘multiple
A modification of the SPI model was later proposed by memory systems’’ here simply to emphasize that different types
Graham et al. (2000). They proposed that there is a direct of memory (episodic/semantic or recollection/familiarity) arise at
different stages of interaction within a hierarchy of brain regions
pathway between perceptual systems and the episodic system (a better name for PIMMS might be ‘‘predictive interactive mul-
that ‘‘bypasses’’ the semantic system. They argued that this tiple memory signals’’). Note also that we do not further charac-
pathway is necessary to explain why recognition memory can terize these memory systems in terms of different states of con-
be spared in individuals with Semantic Dementia, despite sciousness, as does Tulving (2002). We suggest instead that the
their compromised semantic system, particularly for percep- nature of conscious states—e.g., explicit vs. implicit memory
(Graf and Schacter, 1985)—is an orthogonal issue (see also Slot-
tually-rich but less meaningful stimuli. This pathway may nick, 2010), which most likely depends on interactions between
also explain how recollection could be intact despite impaired these memory systems and other neural systems, e.g., prefron-
familiarity, e.g., following a perirhinal lesion (Bowles et al., tal-parietal networks, Dehaene et al. (2006).

Hippocampus
 PREDICTIVE, INTERACTIVE MULTIPLE MEMORY SYSTEMS 1319

episodic system is to bind items to their episodic context. By

‘‘item,’’ we refer to visual objects or auditory tokens that are
the focus of attention (entailed, for example, by the nature of a
memory task). More precisely, an ‘‘item’’ is a hypothetical cause
of the current ‘‘stimulus’’ that impinges on the senses. By
‘‘context,’’ we refer loosely to other background information
like the spatial environment, one’s internal thoughts and emo-
tions, and possibly some form of temporal information. We
assume the episodic system is also necessary to encode single
events defined by the cooccurrence of two or more unrelated
items, for example in associative priming (e.g., Degonda et al.,
2005). The primary purpose of the semantic system, on the
FIGURE 3. The relationship between brain, hypothetical mem-
other hand, is to store combinations of perceptual features that ory system, and behavior in PIMMS. Routes a–f reflect different
repeatedly cooccur (in certain relationships) in the environ- possible causes of behavioral outcomes in recognition memory or
ment, and which thus define items (e.g., Murray et al., 2007; perceptual priming paradigms (see text).
Cowell et al., 2010). This includes binding together features
that are represented within different perceptual systems (sensory
modalities), and also ‘‘chunking’’ recurring sets of items into
larger item representations. The role of the perceptual systems tions and the current ‘‘bottom-up’’ input. The purpose of the
is to abstract and represent those recurring features from the recurrent interactions throughout the systems is then to mini-
environment that define items. mize this so-called ‘‘prediction error’’ (Mumford, 1992; Dayan
To make contact with the neuroimaging data, our second and Hinton, 1996; Rao and Ballard, 1999). For example, the
assumption is the hippocampus is a key component of the activity of possible representations of the currently-attended
episodic system, the perirhinal cortex is a key component of visual object in perirhinal cortex is adjusted on the basis of the
the semantic system (though extending to anterior temporal prediction error passed forward from activity reflecting possible
lobes more generally, Patterson et al., 2007) and the more features in occipitotemporal cortex. Perception of a specific
posterior cortices that are specialized for a particular sensory object (after a few hundreds of milliseconds of such recurrent
modality comprise the perceptual systems (e.g., the ventral processing) occurs when the network of systems settles into a
visual pathway in occipitotemporal cortex or the auditory stable state in which the overall prediction error has been
pathway in lateral temporal cortex). Note that this anatomical minimized.
mapping is tentative (see Fig. 3), and each system may of While the recurrent flow of activity that minimizes predic-
course include several other interconnected regions (e.g., tion error reflects the process of perception, memory arises
Aggleton and Brown, 1999), not to mention many brain from subsequent changes in synaptic connections between
regions important for other factors that affect memory, like systems. The nature of these synaptic changes is again such
motivation, attention, emotionality, and organization, but that they reduce prediction error that arises when an item/con-
which are not considered here. Our conception thus shares text recurs (by maximizing the ‘‘free energy’’ of the network
many similarities with other recent functional-neuroanatomi- as a whole, Friston, 2010). Thus larger residual prediction
cal models of memory (e.g., Davachi, 2006; Diana et al., errors generally entail greater memory storage. Note that the
2007; Eichenbaum et al., 2007; Ranganath, 2010). However, synaptic change necessary for long-term storage is assumed to
our novel addition is that there is not only feedforward of in- occur ‘‘offline,’’ i.e., possibly minutes to days after encoding a
formation during encoding (as in the SPI model), but also stimulus.
feedback of information, during both encoding and retrieval. This extension of the neurophysiological/computational
The important consequence is that encoding and retrieval are idea of ‘‘predictive coding’’ to higher-level memory systems,
no longer separate processes within each memory system, but particularly contextual predictions from episodic to semantic
arise from recurrent interactions between all three systems memory systems, is why we call this framework ‘‘predictive
(see later). interactive multiple memory systems’’ (PIMMS). Furthermore,
Thus we have implicitly defined a hierarchy, with the hippo- by basing this framework on the inherent tendency of the
campal episodic system at its apex (e.g., Mishkin et al., 1998). brain to attempt to predict its surrounding environment,
The role of feedback from one system is to predict the activity PIMMS represents a theory of human memory that relates the
in ‘‘lower’’ systems in this hierarchy. For example, a representa- psychological and neural causes of encoding and retrieval
tion of the current context (e.g., of the spatial environment, processes to a single principle—the minimization of predic-
perhaps represented in parahippocampal regions) is used by the tion-error—while maintaining distinctions between different
hippocampus to predict items that are likely to appear in that types of representation that distinguish traditional multiple
context, i.e., to predict activity in the perirhinal cortex. The memory systems.
role of the feedforward flow of information, on the other hand, PIMMS is not the first attempt to characterize neural/cogni-
is to transmit the difference between such ‘‘top-down’’ predic- tive processes within a predictive framework (e.g., Bar, 2007;

Hippocampus
1320 HENSON AND GAGNEPAIN

Buckner, 2010; though see also Bubic et al., 2010). However, earlier, are actually rare, though see Richardson-Klavehn et al.,
we do not use ‘‘prediction’’ here in a general sense, e.g., to 1999).ii
refer to anticipation/simulation of future events (cf. Buckner,
2010). Though such temporal predictions may be an impor-
tant function of episodic memory and the MTL, and a wor- fMRI Correlates of Encoding and Retrieval
thy goal for extensions of PIMMS, our use here is restricted Related to Familiarity and Recollection
to predictions about the nature of the current stimulus (per- According to PIMMS
ception) on the basis of past experience (memory): more pre-
cisely, predictions about one type of information (e.g., of The ‘‘on-line’’ stabilization process that occurs due to recur-
item representations in perirhinal cortex) from a different rent minimization of prediction error over the timescale of
type of information in a higher layer (e.g., of context repre- milliseconds (e.g., after presentation of a recognition test cue)
sentations in hippocampus). PIMMS is therefore the first is equivalent to the process of ‘‘retrieval’’; the ‘‘off-line’’ minimi-
attempt to (1) describe how predictions can be framed within zation that occurs over the longer timescale due to synaptic
an interactive, multiple memory system view, and (2) relate change is equivalent to learning/storage. This ‘‘storage’’ stage is
these to behavioral and neuroimaging data on recognition not easily detectable with the typical ‘‘online’’ measures of func-
memory. tional neuroimaging. However, the precursors of storage—the
process of memory encoding—can be examined by comparing
the online activity during presentation of a stimulus as a function
of whether that stimulus is later remembered, so-called ‘‘subse-
Behavioral Outcomes According to PIMMS quent memory’’ effects (Paller and Wagner, 2002). According to
The neural determinants of behavioral phenomena depend PIMMS, the increased fMRI signal often observed for such
on the specific task demands (presumably implemented by subsequent memory effects is expected because larger residual
decision mechanisms in frontal/parietal cortices that ‘‘read-out’’ prediction errors (after perception/retrieval has occurred, i.e.,
the neural activity from the relevant brain region). Thus in when the network has stabilized) entail greater synaptic change,
tasks emphasizing visual perceptual decisions, priming will be which will tend to lead to more successful encoding. Thus
driven by ‘‘read-out’’ of memory-related changes in activity in according to PIMMS (and unlike SPI), encoding and retrieval
occipitotemporal cortex, even if the level of such activity is still are intimately related in the sense that they both depend on
affected the feedback of activity from perirhinal and/or hippo- the stabilization of the state of all memory systems into one in
campal regions. In tasks emphasizing contextual retrieval, on which global prediction error is minimized on the basis of
the other hand, recollection should be driven by read-out of stored information.
multiple regions, including hippocampus, perirhinal cortex and In general, prediction errors will be greatest when an unex-
possibly occipitotemporal cortex. pected stimulus is encountered; such as a novel item (leading
However, the interactive nature of PIMMS makes it more to large prediction error in perceptual systems, e.g., occipito-
difficult to isolate different memory systems on the basis of temporal cortex, based on an inability of the semantic system
behavioral data alone. This is because it is difficult to distin- to predict the activity there) or a familiar item in a novel
guish ‘‘contamination’’ of a memory test that is supposed to context (leading to large prediction error in perirhinal cortex,
depend on only one memory system (a ‘‘process-pure’’ test, based on an inability of the episodic system to predict the
Jacoby, 1992) from true neural interactions between two or activity there). Such novelty is accepted as an important deter-
more memory systems. Thus the recognition memory task may minant of memory formation, in that it makes evolutionary/
be an impure test of episodic memory either because it can be metabolic sense not to repeatedly encode information that is
solved by independent contributions from recollection and already well-predicted by the brain. Below, we focus on the
familiarity (as assumed by the SPI and dual-process models more specific fMRI predictions of PIMMS for encoding and
described earlier)—routes b and c in Figure 3—or because retrieval related to familiarity and recollection in the recogni-
‘‘retrieval’’ of contextual information from the episodic memory tion memory paradigm.
system depends on interactions with the semantic memory
system (or perceptual systems)—routes a and b in Figure 3.
ii
Likewise, a test of repetition priming might include a contribu- One empirical observation often taken as evidence of inde-
tion from a perceptual system, in addition to independent pendent retrieval from multiple memory systems (as in SPI) is
‘‘stochastic independence’’: the failure to find significant condi-
‘‘contamination’’ by retrieval from the episodic system—routes tional probability between performance in two memory tests on
e and f in Figure 3—or true interactions between perceptual the same stimulus. However, testing for such stochastic inde-
and episodic systems—routes d and f in Figure 3. This high pendence requires some care (Poldrack, 1996), and the addition
level of interaction may also explain why most dissociations in of nonmnemonic, but task-specific noise processes can dramati-
behavioral performance across different memory tasks are single cally reduce the degree of dependence predicted by single-sys-
tem models (Berry et al., 2008). Moreover, stochastic independ-
dissociations, which are easily explained by scaling effects and/ ence has been found even between two cues within the same
or task-specific noise processes (e.g., Berry et al., 2008), (and memory test (Hayman and Tulving, 1989), questioning this mea-
why the more compelling reversed associations, as defined sure as an indicator of independent memory systems.

Hippocampus
 PREDICTIVE, INTERACTIVE MULTIPLE MEMORY SYSTEMS 1321

Given its slow dynamics, the fMRI signal reflects the integra- The mechanisms of priming and familiarity are therefore
tion of neural activity over time, and so PIMMS predicts related, in that both depend on synaptic changes between the
higher fMRI signals when prediction error is greatest, and the semantic system and a perceptual system (differing mainly as a
network tends to take longer to settle into a stable state.iii The consequence of synaptic change from the semantic to the per-
reduction in the integrated prediction error over the few hun- ceptual system in the case of priming, or from the perceptual
dred milliseconds after stimulus onset has already been pro- to the semantic system in the case of familiarity). This is con-
posed as the neural mechanism of ‘‘repetition suppression’’ sistent with claims that priming and familiarity derive from the
(Henson, 2003). Repetition suppression refers to the reduced same cause—processing fluency (Jacoby and Dallas, 1981)—
fMRI response for repeated vs. initial presentations of a stimu- perhaps differing in practice according whether the task orients
lus, and is sometimes accompanied by an earlier response la- the participant toward perceptual fluency (prediction error in
tency (Gagnepain et al., 2008a), as predicted by a shorter dura- occipitotemporal cortex) or conceptual fluency (prediction error
tion of underlying neural activity (Henson et al., 2002). in perirhinal cortex). However the psychological concept of a
According to PIMMS, repetition suppression in, for example, ‘‘feeling of familiarity’’ is clearly more complex than such sim-
occipitotemporal cortex can be interpreted in terms of the cur- ple distinctions, since feelings of familiarity only arise when
rent visual input being more rapidly ‘‘explained away’’ because processing fluency is unexpected (e.g., one might process one’s
of improved predictions for the perceptual features associated toothbrush fluently every evening, but this does not often lead
with (recently seen) items; improvements caused by synaptic to a feeling of familiarity attributed to the past, Whittlesea and
changes triggered by the prediction error that occurred after Williams, 1998). This concept of familiarity should not
the initial presentation of the stimulus. therefore be directly equated with prediction error in perirhinal
Repetition suppression in occipitotemporal brain regions has cortex; rather prediction error corresponds to fluency, whereas
been associated with perceptual priming. However, repetition familiarity is the attribution of that fluency to the past. A full
suppression is also found in more anterior temporal regions theory of the concept of familiarity is thus beyond current
like perirhinal cortex (Henson et al., 2003), which has been considerations, and needs to include meta-memory factors such
traditionally associated with familiarity instead.iv Assuming an as the expected level of fluency in a certain situation.
absence of strong predictions about a studied or unstudied In the case of recollection, PIMMS states that a key function
item based on the current context (e.g., via feedback from of the hippocampus is to optimize the mutual predictability
hippocampus)—and absence of a strong prediction about the between items (represented in perirhinal cortex) and contexts
context from the studied item (see below)—then the only (presumably represented in multiple regions depending on the
difference between studied and unstudied items at test will be type of context, though here we will focus on spatial context,
the smaller prediction error fed forward from perceptual and assume this is represented in parahippocampal cortex).
systems because of synaptic change after study. Therefore a Mutual predictability corresponds to the joint probability of
smaller fMRI response is predicted in perirhinal cortex for predicting an item from a context and predicting a context
studied than unstudied items. Note that this greater input to from an item. For example, even though a context may not
perirhinal cortex for unstudied items, in conjunction with little strongly predict a particular item, an item may predict a differ-
difference in contextual predictions fed back from hippocampus ent context, leading to high mutual prediction error in the hip-
for studied or unstudied items, in turn entails a greater predic- pocampus. It is this hippocampal prediction error that drives
tion error fed forward to hippocampus, explaining why CRs episodic encoding, i.e., synaptic change between hippocampus
can increase fMRI signal relative to C2 trials in the hippocam- and perirhinal (and parahippocampal) cortices. Therefore those
pus (Fig. 1C). Indeed, the recurrent activity between perirhinal items that lead to high prediction error in hippocampus (for
cortex and hippocampus may entail greater functional coupling reasons including the precision of their item representation and
between these regions (Fernandez and Tendolkar, 2006), in the strength of their item-context associations; see ‘‘butcher in
turn leading to encoding of CRs with the test context (Buckner office’’ example below) will tend to produce greater hippocam-
et al., 2001; Stark and Okado, 2003). pus fMRI signal in the study phase of a recognition memory
experiment that is associated with subsequent recollection in
the test phase.
Later during the test phase, these study-induced synaptic
iii
changes mean that the re-presentation of that item (leading to
The dynamics of the resolution of prediction error may be visi- a pattern of activity in perirhinal cortex) produces a better
ble with other techniques like EEG and MEG—indeed the parie-
tal variant of the P600 that has been associated with recollec- reinstantiation of a representation of its study context. How-
tion might relate to prediction error in the episodic system, ever, because the current context of the recognition test phase
while the earlier frontal variant of the N400 that has been asso- does not match the reinstantiated context, this produces greater
ciated with familiarity might relate to prediction error in the prediction error in the hippocampus (and possibly the para-
semantic system (see Rugg and Curran, 2007; Paller et al., hippocampal cortex)—i.e., greater fMRI signal during recollec-
2007)—but such considerations are beyond the present remit.
iv
Note that reduced activity to repeated stimuli can be found in tion. Conversely, if no context is predicted by the item (i.e., no
other brain regions too, including the hippocampus, for example study context is reinstantiated), then there is little prediction
in single cell responses to scenes (Yanike et al., 2009). error in the hippocampus to minimize, hence little fMRI signal

Hippocampus
1322 HENSON AND GAGNEPAIN

tem above. The difference between the two density functions

determines the initial prediction error. After settling into a
steady state in which this prediction error is minimized, the
activity in the semantic system, for example, can be interpreted
as the posterior probability (the dashed lines in Fig. 4) of each
item being present, given the contextual priors and perceptual
evidence. The difference between the posterior and prior
densities after the network has stabilized is then the residual
prediction error that drives subsequent learning (such that the
future prior densities are moved slightly toward the posterior
density). Thus a high residual prediction error received by the
semantic system from the perceptual system, for example, will
lead to semantic encoding, while a high residual prediction
error received by the episodic system from the semantic or
perceptual systems will lead to episodic encoding.
FIGURE 4. Illustration of prior and likelihood probability The prior distributions in Figure 4 can have either low
density functions within a layer of topographically organized or high variance, corresponding to either precise (‘‘strong’’)
neurons comprising a memory system. The x-axis might reflect, for or imprecise (‘‘weak’’) predictions, respectively. The evidence
example, the features in a perceptual memory system, or items in a
semantic memory system, organized by similarity along a single (likelihood) can be similarly precise or imprecise. In the
dimension for illustrative purposes (in reality the distributions extreme case where there are no predictions (for example in a
might be multidimensional). The prior, from a higher layer in completely novel context), the priors are ‘‘flat’’ (e.g., any item
PIMMS, is shown in heavy lines; the data likelihood (or is deemed equally likely to occur). In addition, a prior can also
‘‘evidence’’), from a lower layer, is shown in light lines (and the
be more or less accurate (depending on how close its central
posterior probability is shown by the dashed lines). The integrated
difference between prior and evidence densities gives the initial tendency is to the central tendency of evidence). Thus an
prediction error (PE) shown in each panel. The top row shows ‘‘accurate’’ contextual prior means that an expected item is pres-
cases with fixed evidence but different prior precisions; specifically: ent, whereas an inaccurate prior means that an item different
a precise and accurate prior (Panel A, e.g., ‘‘butcher in the butch- from expected is present. Assuming that the fMRI signal not
er’s shop’’); an imprecise prior (Panel B, e.g., ‘‘butcher on the
only integrates neural activity over the time taken to minimize
bus’’); or a precise but inaccurate prior (Panel C, e.g., ‘‘butcher in
one’s office’’). Panels D and E show the converse case of an impre- prediction error, but also integrates activity spatially across all
cise (flat) prior together with either imprecise evidence (e.g., a neurons in the layer in Figure 4, then PIMMS predicts that
meaningless stimulus) or more precise evidence (e.g., a primed the fMRI signal will tend to increase as the integrated differ-
stimulus) respectively. Panel F shows the case for low-frequency ence (divergence) between the prior and likelihood densities
items, where the evidence is less precise, but the prior is biased
increases.v We now show how considering priors and evidence
toward other (high-frequency) items.
in terms of both their precision and accuracy can help make
testable predictions for everyday situations.
there. In a simple toy example in the Supporting Information,
we give a concrete example of how a simple two-layer version
Everyday Memory and New Experimental
of PIMMS that distinguishes between these context-to-item
Predictions From PIMMS
associations and item-to-context associations can explain the The importance of contextual priors:
qualitatively different patterns of fMRI signal in Figure 1C. ‘‘The butcher in the office’’
Below, we expand the above accounts in terms of probability
A probabilistic conception of PIMMS suggests some predic-
densities.
tions for the fMRI signal in perirhinal cortex (semantic system)
as a function of three different types of context: (1) a precise
A Probabilistic Perspective on PIMMS and accurate prior (Fig. 4A), (2) an imprecise prior (Fig. 4B),
and (3) a precise but inaccurate prior (Fig. 4C). Because the
One way to think about PIMMS is in terms of prior and prediction error increases across these three contexts (i.e.,
likelihood probability densities. For illustrative purposes, greater divergence between prior and likelihood distributions),
imagine a single layer of neurons with tuning curves that are the fMRI signal is expected to increase in perirhinal cortex.
topographically organized, such that similar contexts/items/fea-
tures are represented close together. Probability densities reflect-
v
ing interpretation of an input to that layer are then likely to be In general however, note that neural activity in PIMMS does
unimodal distributions (Fig. 4). The likelihood density (the not need to conform to probability densities (i.e., it is not con-
strained to sum to one), in that total integrated activity within a
light lines in Fig. 4) represents the data, or ‘‘evidence,’’ coming layer of neurons can change (e.g., due to ‘‘global’’ attentional
from the system below, and the prior density (the heavy lines effects), without any necessary change in the accuracy or preci-
in Fig. 4) corresponds to the prediction coming from the sys- sion of the priors/evidence.

Hippocampus
 PREDICTIVE, INTERACTIVE MULTIPLE MEMORY SYSTEMS 1323

This theoretical prediction may not be very surprising, but the encountered (unlike any known item), then there will be a flat
assumption of PIMMS that the amount of subsequent synaptic distribution of activity across neurons in the semantic system.
change between perirhinal and hippocampal systems is deter- This is because, even if there is reasonably precise input to the
mined by the amount of residual prediction error leads to the perceptual system, the lack of any unique predictions for
further experimental prediction that the occurrence of an item that combination of features (in the item-to-feature and fea-
that is not expected in a context (Fig. 4C) will be better ture-to-item connections) results in an imprecise likelihood
episodically encoded than occurrence of the same item in a within the semantic system. The prediction error forwarded to
context that has no strong expectancy (Fig. 4B). the episodic system between a flat (episodic) prior and such flat
An everyday example of this prediction would be a modifica- evidence (Fig. 4D) is minimal (i.e., there is little information that
tion of the famous ‘‘butcher on the bus’’ example. This exam- can be learned). In turn, episodic encoding of such unknown
ple is normally used to illustrate familiarity—i.e., a sense of items (in terms of weight changes between semantic and episodic
knowing the person but not remembering who they are—but systems) is minimal. This would be analogous to the standard
the question here is how well this event would be encoded ‘‘butcher on the bus’’ situation: if the butcher is not identified as
(and in the examples below, we are assuming that the butcher such (the posterior distribution over the item layer remains
is, in fact, recognized as the butcher). As a type of public trans- imprecise), then that situation will not be well-encoded.vii If a
port, one could argue that the context of being on a bus does stimulus is better known (or more closely resembles known
not impose strong predictions about the people one might items), on the other hand, there will be a more precise data like-
encounter there. In the context of the butcher’s shop however, lihood, greater prediction error, and hence greater episodic encod-
there is a strong prior for the butcher, while in one’s office at ing (Fig. 4E). This can explain the better recognition memory
work, there are strong priors for one’s work colleagues instead.vi for novel stimuli (e.g., visual squiggles) that are found to be
According to the above argument, PIMMS predicts that one more ‘‘meaningful’’ (Voss and Paller, 2007).
will encode the event of ‘‘butcher in the office’’ better than This explanation may seem difficult to reconcile with
‘‘butcher on the bus’’ (and both should of course be better evidence that recollection is often better for low than high
encoded than the fully-expected ‘‘butcher in the butcher’s frequency words (for review, Yonelinas, 2002), but in fact,
shop’’). In other words, if one later encountered the butcher in PIMMS can predict greater episodic encoding for low than
a new town (another context with weak priors); the prediction high frequency items if the contextual prior is assumed to be
is that one would be more likely to recollect his previous imprecise, but not flat—specifically, slightly biased toward high
appearance in one’s office than his previous appearance on a frequency items (Fig. 4F). While the evidence (input from
bus. Testing the above prediction of PIMMS in the laboratory perceptual systems) may be less precise for low- than for high-
(i.e., fMRI scanner) requires imposing specific contexts and frequency items, the presence of some information in the
varying the precision of associations between those contexts feature-to-item associations from the perceptual layer (unlike
and specific items. We are not aware of such an fMRI experi- the above case of a truly unknown stimulus) means that the
ment to date, but would predict that such an experiment data likelihood for a low frequency-item is not flat. Impor-
would show both increased perirhinal and hippocampal activity, tantly, its central tendency will be further removed from that of
and increased perirhinal-hippocampal connectivity (see below), the prior than would be the central tendency of a high-frequency
for those conditions analogous to the ‘‘butcher in the office,’’ item, which can lead to a greater relative prediction error for
relative to those analogous to the ‘‘butcher on the bus.’’ low-frequency items and hence a greater updating of item-
context connections (i.e., greater episodic encoding; though
effects of frequency at encoding are also likely to involve other
The importance of item priors: factors such as attention; Diana and Reder, 2006).
‘‘The recently seen butcher’’ Lastly, consider for example the effect of repetition priming
on episodic encoding. Take an extended recognition memory
The perspective of precision and accuracy can also help
explain the role of prior item knowledge in recognition mem-
ory. Consider the case when the (episodic) prior on the seman- vii
tic system is flat—i.e., no particular items are predicted in Note that we do not distinguish here between a representation
of a familiar face, and a representation of the identity of that
the current context—but the evidence from lower perceptual face (both are subsumed within the semantic system here). In a
systems varies (Figs. 4D,E). If a completely novel stimulus is more complex model with yet further levels of representation, it
might be possible to explain how a familiar face (the unrecog-
nised butcher) can be well encoded (i.e., later recollected) as
vi
Note that these ‘‘semanticized’’ associations between a context having occurred in a specific context (owing to plasticity
(e.g., butcher’s shop) and an item (e.g., butcher) may in fact be between that layer and a contextual layer, e.g., hippocampus),
instantiated in direct connections between parahippocampal even if that face is never identified as the butcher’s. We are also
and perirhinal cortex, rather than via the hippocampus (an issue aware that we do not address the issue of how new semantic
we will explore in extensions of PIMMS), but for the purpose of representations of completely novel stimuli are ever learned,
the predictions for the laboratory (fMRI) experiment below, we which is a difficult question that we leave for further develop-
assume that experimentally-established context-item associa- ments of PIMMS, though is likely to entail changes in lateral
tions engage the hippocampus, i.e., the episodic system. connections within a region.

Hippocampus
1324 HENSON AND GAGNEPAIN

paradigm in which some of the items in the study phase are generally expected when the prediction error in both regions is
pre-exposed in a preceding ‘‘priming phase.’’ One effect of this high (in that greater interaction between regions is needed to
prior perceptual processing of a stimulus is to ‘‘sharpen’’ the reduce that error), but can also occur when prediction error is
feature-item connections, leading to a more precise distribution rapidly reduced in one region because of activity in the other:
of evidential input into the semantic system. In conjunction i.e., effective connectivity from Region A to Region B is high
with an uninformative contextual prior, this sharper likelihood when activity changes in Region B are highly dependent on ac-
produces greater prediction error and in turn greater episodic tivity in Region A. PIMMS predicts that familiarity will be
encoding (cf. Figs. 4B and E). Assuming that the context in associated with changes in the effective connectivity between
most laboratory experiments conforms to such a weak prior semantic and perceptual systems, while recollection will be
(i.e., though a participant may become aware that only words associated with increased effective connectivity between epi-
or pictures will be presented, the probability of any particular sodic and semantic and/or perceptual systems (Fig. 2B). Sup-
exemplar of that category is normally fairly constant), this port for the latter hypothesis comes from a recent fMRI
improved episodic encoding for perceptually primed relative to experiment that examined subsequent memory effects for
unprimed words is exactly what was found in a recent behav- primed and unprimed stimuli, using the extended recognition
ioral experiment (Gagnepain et al., 2008b; see also Poppenk paradigm described above (Gagnepain et al., in press). In this
et al., 2010). experiment, the result of complex interactions between episodic
Furthermore, the primed items are likely to have addition- and perceptual systems like those described above was that the
ally been associated with a different context (that of the neural correlates of encoding that led to subsequent recollec-
priming phase), which might be reinstated when they are tion were not apparent in the local pattern of fMRI signal in
presented again during the main study phase. According to the hippocampus per se, but rather in increased effective con-
PIMMS, this will in turn increase the mutual prediction error nectivity between the hippocampus and a brain region (in the
in the hippocampus, again affecting episodic encoding of the lateral, superior temporal cortex) that was associated with the
new study context. Returning to the above ‘‘butcher in the perceptual priming of the auditory words.
office’’ example, the strength of encoding in episodic memory
will depend not only whether or not one has seen the butcher
recently (i.e., whether his face has been primed), but also
whether those occasions occurred in one or a varied set of CONCLUSIONS
contexts. This trade-off between effects at perceptual, semantic
and episodic levels makes the precise patterns of local fMRI We have outlined a theoretical framework—PIMMS—and
signal in each brain region difficult to predict uniquely. This is applied it to the topic of this special issue, viz. the neural
the kind of situation where the effective connectivity between correlates of recollection and familiarity. We have focused on
brain regions might be better able to dissociate familiarity and evidence from fMRI on healthy human volunteers during
recollection. recognition memory paradigms. However, we believe PIMMS
can be extended to many other phenomena in memory (and
perception), such as priming and recall paradigms, and data
The importance of effective connectivity from methods such as lesion studies and single-cell recording,
As described above, PIMMS proposes that recollection and including data from brain regions outside the MTL.
familiarity depend on complex spatial and temporal interac- While we have used the phrase ‘‘memory systems’’ to relate
tions between all three memory systems. Coupled with the PIMMS to traditional theories of memory, we believe predic-
temporal and spatial integration of the fMRI signal, this means tion-error-driven plasticity is a general property of brain
that, while local patterns of fMRI signal within brain regions regions, supporting both memory and perception within those
may be predicted in certain situations, it will be difficult to same regions. Indeed, memory encoding (storage) and percep-
predict this pattern a priori (or uniquely) in many other situa- tion (memory retrieval) are intimately-related in terms of ‘‘off-
tions. Nonetheless, another prediction made by PIMMS is that line’’ vs. ‘‘on-line’’ reduction of prediction error, respectively. At
the pattern of ‘‘effective connectivity’’ between brain regions the same time, we accept that a primary reason why different
should differ during recollection and familiarity. ‘‘Effective brain regions might support different types of memory is
connectivity’’ refers to changes in the functional coupling computational (in the sense of Marr, 1971): in that it seems
between regions as a consequence of an experimental manipu- advantageous, for example, to separate one system that stores
lation.viii Increased functional coupling between two regions is what is unique about experiences (e.g., individuation/pattern
separation)—an episodic system—from another whose primary
function is to extract regularities across experiences (e.g., gener-
viii
More precisely, effective connectivity refers to the causal alization/pattern completion, McClelland et al., 1995)—a
influence that activity in one region has on that in another semantic system. Our main point is that in most everyday
(Friston, 2002), as defined for example by an explicit network
model, and can be framed dynamically in terms of how the rate situations, it also makes sense for these systems to operate
of change of activity in one region depends on the level of ac- together (e.g., to both individuate an event and adjust prior
tivity in another. knowledge about the content of that event). This means that,

Hippocampus
 PREDICTIVE, INTERACTIVE MULTIPLE MEMORY SYSTEMS 1325

unless one or more memory systems are inactive (e.g., from a Fernandez G, Tendolkar I. 2006. The rhinal cortex: ‘‘Gatekeeper’’ of
brain lesion), ‘‘on-line’’ or ‘‘performance’’ measures like behav- the declarative memory system. Trends Cogn Sci 10:358–362.
Friston KJ. 2002. Functional integration and inference in the brain.
ioral outcomes or fMRI will reflect a complex interplay between
Prog Neurobiol 68:113–143.
each system. Friston KJ. 2010. The free-energy principle: A unified brain theory?
Nat Rev Neurosci 11:127–138.
Acknowledgments Gagnepain P, Chetelat G, Landeau B, Dayan J, Eustache F, Lebreton
K. 2008a. Spoken word memory traces within the human auditory
The authors thank Tim Shallice and two anonymous cortex revealed by repetition priming and functional magnetic reso-
reviewers for their helpful comments. nance imaging. J Neurosci 28:5281–5289.
Gagnepain P, Lebreton K, Desgranges B, Eustache F. 2008b. Percep-
tual priming enhances the creation of new episodic memories.
Consc Cogn 17:276–287.
REFERENCES Gagnepain P, Henson R, Chetelat G, Desgranges B, Lebreton K, Eus-
tache F. Is neocortical-hippocampal connectivity a better predicator
Aggleton JP, Brown MW. 1999. Episodic memory, amnesia, and of subsequent recollection than local increases in hippocampal
the hippocampal anterior thalamic axis. Behav Brain Sci 22:425– activity? New insights on the role of priming. J Cogn Neurosci (in
444. press).
Bar M. 2007. The proactive brain: Using analogies and associations to Graf P, Schacter DL. 1985. Implicit and explicit memory for new
generate predictions. Trends Cogn Sci 11:280–289. associations in normal and amnesic subjects. J Exp Psychol Learn
Berry CJ, Shanks DR, Henson RN. 2008. A unitary signal-detection Mem Cogn 11:501–518.
model of implicit and explicit memory. Trends Cogn Sci 12:367– Graham KS, Simons JS, Pratt KH, Patterson K, Hodges JR. 2000.
373. Insights from semantic dementia on the relationship between
Bowles B, Crupi C, Mirsattari SM, Pigott S, Parrent AG, Pruessner episodic and semantic memory. Neuropsychologia 38:313–324.
JC, Yonelinas AP, Kohler S. 2007. Impaired familiarity with pre- Green DM, Swets JA. 1966. Signal Detection Theory and Psychophy-
served recollection after anterior temporal-lobe resection that spares sics. New York: Robert E. Kreiger.
hippocampus. Proc Nat Acad Sci USA 104:16382–16387. Greve A, Donaldson DI, van Rossum MC. 2010. A single-trace dual-
Brown MW, Warburton EC, Aggleton JP. 2010. Recognition memory: process model of episodic memory: A novel computational account
Material, processes and subrtates. Hippocampus 20:1228–1244. of familiarity and recollection. Hippocampus 20:235–251.
Brown MW, Xiang JZ. 1998. Recognition memory: Neuronal sub- Hayman CAG, Tulving E. 1989. Is priming in fragment completion
strates of the judgment of prior occurrence. Prog Neurobiol 55: based on a ‘‘traceless’’ memory system? J Exp Psychol Learn Mem
149–189. Cogn 15:941–956.
Bubic A, von Cramon DY, Schubotz RI. 2010. Prediction, cognition Henson RN. 2003. Neuroimaging studies of priming. Prog Neurobiol
and the brain. Front Hum Neurosci 4:1–15. 70:53–81.
Buckner RL. 2010. The role of the hippocampus in prediction and Henson RN. 2005. What can functional imaging tell the experimental
imagination. Ann Rev Psychol 61:27–48. psychologist? Q J Exp Psychol A 58:193–233.
Buckner RL, Wheeler ME, Sheridan MA. 2001. Encoding processes Henson RN. 2006. Forward inference in functional neuroimaging:
during retrieval tasks. J Cogn Neurosci 13:406–415. Dissociations vs. associations. Trends Cogn Sci 10:64–69.
Carr VA, Rissman J, Wagner AD. 2010. Imaging the human medial Henson RN, Price CJ, Rugg MD, Turner R, Friston KJ. 2002.
temporal lobe with high-resolution fMRI. Neuron 65:298–308. Detecting latency differences in event-related BOLD responses:
Cowell RA, Bussey TJ, Saksida LM. 2010. Components of recognition Application to words versus nonwords and initial versus repeated
memory: Dissociable cognitive processes or just differences in rep- face presentations. Neuroimage 15:83–97.
resentational complexity? Hippocampus 20:1245–1262. Henson RN, Cansino S, Herron JE, Robb WG, Rugg MD. 2003. A
Davachi L. 2006. Item, context and relational episodic encoding in familiarity signal in human anterior medial temporal cortex?
humans. Curr Opin Neurobiol 16:693–700. Hippocampus 13:301–304.
Dayan P, Hinton GE. 1996. Varieties of Helmholtz Machine. Neural Humphreys MS, Pike R, Bain JD, Tehan G. 1989. Global matching:
Networks 9:1385–1403. A comparison of SAM, Minerva II, Matrix and TODAM models.
Degonda N, Mondadori CR, Bosshardt S, Schmidt CF, Boesiger P, J Math Psych 33:36–67.
Nitsch RM, Hock C, Henke K. 2005. Implicit associative learning Jacoby LL. 1992. A process dissociation framework: Separating
engages the hippocampus and interacts with explicit associative automatic from intentional uses of memory. J Mem Lang 30:513–
learning. Neuron 46:505–520. 541.
Dehaene S, Changeux JP, Naccache L, Sackur J, Sergent C. 2006. Con- Jacoby LL, Dallas M. 1981. On the relationship between autobiographi-
scious, preconscious, and subliminal processing: A testable taxonomy. cal memory and perceptual learning. J Exp Psychol Gen 110:306–
Trends Cogn Sci 10:204–211. 340.
Diana RA, Reder LM. 2006. The low-frequency encoding disadvant- Johnson MK, Hashtroudi S, Lindsay DS. 1993. Source monitoring.
age: Word-frequency affects processing demands. J Exp Psychol Learn Psychol Rev 114:3–28.
Mem Cogn 32:805–815. Mandler G. 1980. Recognizing: The judgment pf previous occurence.
Diana RA, Yonelinas AP, Ranganath C. 2007. Imaging recollection Psychol Rev 87:252–271.
and familiarity in the medial temporal lobe: A three-component Marr D. 1971. Simple memory: A theory for archicortex. Phil Trans
model. Trends Cogn Sci 11:379–386. R Soc Lond 262:23–81.
Dunn JC, Kirsner K. 1988. Discovering functionally independent McClelland JL, McNaughton BL, O’Reilly RC. 1995. Why there are
mental processes: The principle of reversed association. Psychol Rev complementary learning systems in the hippocampus and neocor-
95:91–101. tex: Insights from the successes and failures of connectionist models
Eichenbaum H, Yonelinas AP, Ranganath C. 2007. The medial tempo- of learning and memory. Psychol Rev 102:419–457.
ral lobe and recognition memory. Annu Rev Neurosci 30:123–152. Mishkin M, Varga-Khadem F, Gadian D. 1998. Amnesia and the
Eustache F, Desgranges B. 2008. MNESIS: Towards the integration of organization of the hippocampal system. Hippocampus 8:212–
current multisystem models of memory. Neuropsychol Rev 18:53–69. 216.

Hippocampus
1326 HENSON AND GAGNEPAIN

Mumford D. 1992. On the computational architecture of the neocor- Shimamura AP. 2010. Heirarhical relational binding in the medial
tex. II. The role of cortico-cortical loops. Biol Cybern 66:241–251. temporal lobe: The strong get stronger. Hippocampus 20:1206–
Murray EA, Bussey TJ, Saksida LM. 2007. Visual perception and 1216.
memory: A new view of medial temporal lobe function in primates Slotnick SD. 2010. Does the hippocampus mediate objective binding
and rodents. Annu Rev Neurosci 30:99–122. or subjective remembering? Neuroimage 49:1769–1776.
Norman KA. 2010. How hippocampus and cortex contribute to rec- Squire LR, Wixted JT, Clark RE. 2007. Recognition memory and the
ognition memory: Revisiting the complementary learning systems medial temporal lobe: A new perspective. Nat Rev Neurosci 8:
model. Hippocampus 20:1217–1227. 872–883.
Norman KA, O’Reilly RC. 2003. Modeling hippocampal and neocort- Staresina BP, Davachi L. 2008. Selective and shared contributions of
ical contributions to recognition memory: A complementary-learn- the hippocampus and perirhinal cortex to episodic item and asso-
ing-systems approach. Psychol Rev 110:611–646. ciative encoding. J Cogn Neurosci 20:1478–1489.
Paller KA, Wagner AD. 2002. Observing the transformation of experi- Stark CE, Okado Y. 2003. Making memories without trying: Medial
ence into memory. Trends Cogn Sci 6:93–102. temporal lobe activity associated with incidental memory formation
Paller KA, Voss JL, Boehm SG. 2007. Validating neural correlates of during recognition. J Neurosci 23:6748–6753.
familiarity. Trends Cogn Sci 6:243–250. Tulving E. 1985. Memory and consciousness. Can Psychol 26:1–12.
Patterson K, Nestor PJ, Rogers TT. 2007. Where do you know what you Tulving E. 2002. Episodic memory: From mind to brain. Annu Rev
know? The representation of semantic knowledge in the human Psychol 53:1–25.
brain. Nat Rev Neurosci 12:976–987. Tulving E, Gazzaniga MS. 1995. Organization of Memory: Quo Vadis?
Poldrack RA. 1996. On testing for stochastic dissociations. Psychon The Cognitive Neurosciences. Cambridge: MIT Press. pp 839–
Bull Rev 3:434–448. 847.
Poppenk J, McIntosh AR, Craik FI, Moscovitch M. 2010. Past experi- Voss JL, Paller KA. 2007. Neural correlates of conceptual implicit
ence modulates the neural mechanisms of episodic memory forma- memory and their contamination of putative neural correlates of
tion. J Neurosci 30:4707–4716. explicit memory. Learn Mem 14:259–267.
Ranganath C. 2010. A unified framework for the functional organiza- Wais PE, Mickes L, Wixted JT. 2008. Remember/know judgments
tion of the medial temporal lobes and the phenomenology of epi- probe degrees of recollection J Cogn Neurosci 20:400–405.
sodic memory. Hippocampus 20:1263–1290. Whittlesea BWA, Williams LD. 1998. Why do strangers feel familiar,
Rao RP, Ballard DH. 1999. Predictive coding in the visual cortex: A but friends don’t? A discrepancy-attribution account of feelings of
functional interpretation of some extra-classical receptive-field effects. familiarity. Acta Psychol 98:141–165.
Nat Neurosci 2:79–87. Wixted JT, Mickes L, Squire LR. 2010. Measuring recollection and
Richardson-Klavehn A, Clarke AJB, Gardiner JM. 1999. Conjoint dis- familiarity in the medial temporal lobe. Hippocampus 20:1195–
sociations reveal involuntary ‘‘perceptual’’ priming from generating 1205.
at study. Consc Cogn 8:271–284. Yanike M, Wirth S, Smith AC, Brown EN, Suzuki WA. 2009.
Rotello CM, Macmillan NA, Reeder JA. 2004. Sum-difference theory Comparison of associative learning-related signals in the macaque
of remembering and knowing: A two-dimensional signal-detection perirhinal cortex and hippocampus. Cereb Cortex 19:1064–1078.
model. Psychol Rev 111:588–616. Yonelinas AP. 2002. The nature of recollection and familiarity: A
Rugg MD, Curran T. 2007. Event-related potentials and recognition review of 30 years of research. J Mem Lang 46:441–517.
memory. Trends Cogn Sci 11:251–257. Yonelinas AP, Aly M, Wang W-C, Koen JD. 2010. Recollection and
Schacter DL, Tulving E. 1994. Memory Systems 1994. Cambridge: familiarity: Examining controversial assumptions and new direc-
MIT Press. tions. Hippocampus 20:1178–1194.

Hippocampus