doi 10.1098/rstb.2001.

0939

Components of episodic memory:

the contribution of recollection and familiarity

Andrew P. Yonelinas
Department of Psychology, University of California, Davis, CA 95616 USA ([email protected])

The examination of recognition memory con¢dence judgements indicates that there are two separate
components or processes underlying episodic memory. A model that accounts for these results is described
in which a recollection process and a familiarity process are assumed to contribute to recognition
memory performance. Recollection is assumed to re£ect a threshold process whereby qualitative inform-
ation about the study event is retrieved, whereas familiarity re£ects a classical signal-detection process
whereby items exceeding a familiarity response criterion are accepted as having been studied. Evidence
from cognitive, neuropsychological and neuroimaging studies indicate that the model is in agreement
with the existing recognition results, and indicate that recollection and familiarity are behaviourally,
neurally and phenomenologically distinct memory retrieval processes.
Keywords: recognition; familiarity; memory; episodic; recollection;explicit

how items are represented and how these items are stored
1. INTRODUCTION
in the memory. These include global memory models
The notion that episodic memory consists of distinct such as episodic or instance models (e.g. MINERVA,
components dates back at least to Aristotle. In the 1970s Hintzman 1986), as well as connectionist or distributed
and early 1980s, cognitive psychologists formalized this models (e.g. TODAM, Murdock 1982). Although the
notion and developed dual-process models that assumed speci¢c assumptions of these models di¡er, they all main-
that there were two separate processes, recollection and tain that recognition memory judgements rely on the
familiarity, that contributed to episodic recognition assessment of a single familiarity measure.
memory (e.g. Atkinson & Juola 1974; Jacoby & Dallas Over the past 10 years, however, the limitations of the
1981; Mandler 1980; Tulving 1985). The idea was that single process models have become increasingly obvious
previously studied items would be more familiar than (see, for example, Clark & Gronlund 1996; Hockley
new items, thus subjects could accept the more familiar 1991; Ratcli¡ et al. 1992; Yonelinas 1994), and there has
items as having been studied. However, in addition to been a renewed interest in dual-process theories of recog-
assessments of familiarity, if subjects could retrieve some nition memory. The aim of the current paper is to review
aspect of the study event, such as when or where it some of the recognition memory work that my colleagues
occurred, this could also be used as a basis for recogni- and I have conducted over the past 10 years. I will ¢rst
tion judgements. describe a set of ¢ndings that demonstrate that there are
Despite the introspective appeal of the dual-process at least two distinct components of episodic recognition
models and their initial success in accounting for a memory. I will argue that these two components re£ect
variety of behavioural results, the dominant theories of the operation of two distinct retrieval processes: recollec-
that period assumed that recognition memory re£ected tion and familiarity. I will then describe a dual-process
only a single familiarity process, and recollection was not model that was designed to account for these results and
thought to play a signi¢cant, if any, role in recognition review the empirical studies that have been conducted to
memory judgements. In these single process models, test the underlying assumptions of that model. I will
recognition was generally assumed to be well described conclude by discussing the limitations of that model and
by signal-detection theory (see ¢gure 1). The basic idea is raise questions for future studies of episodic memory.
that studied items are on average more familiar than new
items, but because the old and new item familiarity distri-
2. RECOGNITION RECEIVER OPERATING
butions overlap it is necessary to set a response criterion
CHARACTERISTICS
and accept only the items above that level of familiarity
as having been studied. The advantage of the model is One area of research that turns out to be particularly
that it uses only a single memory component, thus recog- problematic for the current single process models of
nition memory accuracy can be characterized using a episodic memory is the study of receiver operating char-
single parameter (i.e. d', which is the distance between acteristics (ROCs). A ROC is the function that relates the
the old and new item distributions). Over the past 20 proportion of correct recognitions (i.e. the hit rate) to the
years, single process models have become more sophisti- proportion of incorrect recognitions (i.e. the false alarm
cated and have included additional assumptions about rate). Typically, performance is examined across levels of

Phil. Trans. R. Soc. Lond. B (2001) 356, 1363^1374 1363 & 2001 The Royal Society
1364 A. P.Yonelinas Components of episodic memory

d' 1

new old 0.8

hits ("old" old)

0.6

response: "new" "old" 0.4

response criterion (c)
familiarity 0.2
Figure 1. Familiarity distributions for old and new items for
an equal-variance signal-detection model.
0
0 0.2 0.4 0.6 0.8 1
false alarms ("old" new)
response con¢dence. For example, after studying a list of
words, subjects are presented with a mixture of old and Figure 2. Symmetrical (lower function) and asymmetrical
new words and are required to make recognition judge- (upper function) receiver operating characteristics.
ments on a scale ranging from `sure it was studied' to `sure
it was not studied'. The ROC is plotted as a function of
con¢dence such that the leftmost point includes only the However, recognition accuracy and the degree of ROC
most con¢dently recognized items and subsequent points asymmetry are functionally independent (e.g. Ratcli¡ et al.
include less and less con¢dent responses. 1992; Glanzer et al. 1999; Yonelinas 1994), indicating that
Figure 2 shows two hypothetical recognition memory recognition memory re£ects at least two separate
ROCs. The lower function is generated by the signal- memory components. That is, in some studies, increases
detection model illustrated in ¢gure 1. The function is in accuracy are accompanied by increases in ROC asym-
produced by plotting the hits against the false alarms as metry, whereas in other studies the degree of ROC asym-
the response criterion is varied. This function is curvi- metry remains relatively constant as accuracy increases.
linear and symmetrical along the diagonal. It is curvi- Thus, there is no way to characterize the existing ROCs
linear because of the continuous nature of the Gaussian using fewer than two separate memory components or
familiarity distributions, and it is symmetrical because parameters.
the old and new item's familiarity distributions are the Using the signal-detection framework, one needs one
same shape (i.e. they have equal variance). component to account for increases in accuracy (i.e. d')
Early memory studies indicated that recognition and another component to account for the changes in
memory ROCs were curvilinear and approximately ROC asymmetry (i.e. the variance of the old item distri-
symmetrical (e.g. Murdock & Dufty 1972). This lent bution relative to the new item variance). This `unequal-
support to the notion that episodic memory re£ected only variance signal-detection model' can produce the data
a single component, and justi¢ed the practice of mea- pattern just described but, as we will see later, it fails to
suring recognition accuracy using a single parameter (e.g. account for the ROCs observed in recognition memory
d', A', or proportion correct). However, subsequent tests.
studies demonstrated that recognition ROCs were not This simple pattern of results turns out to be extremely
generally symmetrical but rather took the form of the top problematic for current single process models, even
function in ¢gure 2. The ROC is curvilinear but it is multiple parameter models such as MINERVA and
asymmetrical, and appears to be pushed up along the left TODAM. The models either predict that the ROC asym-
y-axis. In terms of signal-detection theory, this asym- metry should always remain constant or that it should
metry indicates that the old item familiarity distribution always increase as accuracy increases and thus they
must be associated with more variance (i.e. a fatter distri- cannot account for the fact that both patterns are
bution) than the new item distribution. observed. The problem is that these models do not have
The fact that ROCs are asymmetrical is not necessarily separate parameters that are tied to accuracy and ROC
a problem for single component views of recognition asymmetry and it is not possible to introduce separate
memory. For example, if the degree of asymmetry were components without restructuring these models (see
always the same, then one could assume that the old item Ratcli¡ et al. 1992).
variance was always some constant amount greater than
the new item variance, and thus it would still be possible
3. THE DUAL-PROCESS SIGNAL-DETECTION MODEL
to measure recognition accuracy using a single d' para-
meter. Alternatively, if the degree of asymmetry were The important question arising from the ROC litera-
directly related to accuracy (e.g. as accuracy increased, ture is, `why are these dissociations observed in simple
the degree of asymmetry always increased) then one recognition memory tests?'. The explanation provided by
could still measure recognition performance using a dual-process theory is that the dissociations occur
single accuracy parameter. because there are two retrieval processes, rather than just

Phil. Trans. R. Soc. Lond. B (2001)

 Components of episodic memory A. P.Yonelinas 1365

one, that contribute to recognition performance. That is, represents the probability that a studied item is recol-
recognition memory judgements can be based either on lected, and d', which represents the average increase in
the assessment of familiarity or on a recollection process familiarity associated with studying an item.
whereby subjects retrieve qualitative information about a Given that there are two processes that di¡erentially
study event. I will argue that the familiarity process contribute to the shape of the recognition ROC, the dual-
produces an ROC that is curvilinear and symmetrical, process model can account for the observed dissociations
whereas the recollection process leads the ROC to between accuracy and asymmetry. That is, according to
become asymmetrical. Because the relative contributions the model the asymmetry typically seen in recognition
of recollection and familiarity can vary, accuracy and ROCs re£ects the fact the recollection is contributing to
ROC asymmetry can vary independently. performance. If recollection increases and familiarity
These assumptions form the basis of a simple quanti- remains relatively constant then accuracy should increase
tative model that I will refer to as the dual-process signal- and the ROC should become more asymmetrical. Thus
detection model. The model assumes that familiarity is the model can account for cases in which increases in
well described by the classical signal-detection model accuracy are accompanied by increases in ROC asym-
illustrated in ¢gure 1 (i.e. an equal-variance model). In metry (e.g. Donaldson & Murdock 1968). The model can
contrast, recollection is assumed to re£ect a fundament- also account for cases in which increases in accuracy do
ally di¡erent form of memory retrieval ö a threshold not in£uence the degree of asymmetry (e.g. Ratcli¡ et al.
retrieval process. Describing recollection as a threshold 1992). That is, if recollection and familiarity increase
process means that for any given item a subject either approximately equally then the increase in asymmetry
succeeds at retrieving some information about the study caused by recollection will be o¡set by the increase in
event or they fail to. That is, for some items they may symmetry caused by additional familiarity (see Yonelinas
retrieve information about when or where the item was 1994 for an illustration of these predictions).
presented, but there will be some items that fall below the Further support for the model comes from the ¢nding
threshold, and for these items subjects will be unable to that the shape of the recognition ROC is directly related
retrieve any accurate qualitative information about the to the contribution of recollection and familiarity. For
study event. example, Jacoby's process dissociation procedure (Jacoby
If performance relies exclusively on familiarity then the 1991) was used to estimate the contribution of recollection
model predicts a curvilinear ROC that is symmetrical and familiarity, in order to determine the relationship
along the diagonal (e.g. the lower function in ¢gure 2 between the shape of the ROC and the contribution of
that is generated by an equal variance signal-detection these two processes (Yonelinas 1994). Subjects were
model). If subjects recollect some proportion of the required to make both recognition con¢dence judgements,
studied items then this will increase the hit rate and in£u- and list discrimination judgements indicating from which
ence the shape of the ROC. However, in order to know of two study lists the test items originated. The con¢dence
exactly how it will in£uence the ROC it is necessary to responses were used to plot ROCs. Recollection was then
make assumptions about how recollection and familiarity estimated as the ability to determine list membership
combine. The model assumes that the two processes make accurately and familiarity was estimated as the prob-
independent contributions to recognition, and that recol- ability of recognizing an item, given that it was not accu-
lection leads to relatively high con¢dence recognition rately recollected. As expected, the results across several
responses. Thus recollection will add high con¢dence hits, experiments showed that when recollection increased but
and the leftmost point on the ROC will move up. Because familiarity was unchanged, accuracy increased while the
the ROC is cumulative across con¢dence, the entire ROC ROCs became more asymmetrical. Moreover, when both
will be shifted up and thus become asymmetrical (e.g. the recollection and familiarity increased together, accuracy
top function in ¢gure 2). increased while the ROC asymmetry remained constant.
The model can be represented by the following Finally, estimates of recollection and familiarity derived
equations: from the process dissociation procedure were found to
predict the observed recognition con¢dence ROCs accu-
P(`old'jold) ˆ R ‡ (1 R)Fo (3:1) rately. The results indicate that the shape of recognition
P(`old'jnew) ˆ Fn (3:2). ROCs is directly related to the contribution of recollec-
tion and familiarity.
Old items will be correctly recognized if they are recol- The dual-process model can therefore account for the
lected (R), or if they are familiar (Fo) in the absence of existing recognition memory ROC results that are problem-
recollection (17R). New items will be incorrectly atic for earlier models, and it shows that the shape of the
accepted as old if they are familiar (Fn). If familiarity is recognition ROC is directly related to the contribution of
assumed to re£ect a signal-detection process then Fo and recollection and familiarity. Although these results
Fn will be a function of d' (the distance between the provide support for the model, one would like to be able
means of the old and new item distributions) and c (the to test the individual assumptions underlying the model
response criterion), such that Fo ˆ (d'/27c) and directly. One advantage of the model is that it is based on a
Fn ˆ (7d'/27c). These functions represent the propor- relatively small number of assumptions, and thus it is
tion of the old and new item distributions that exceed the possible to assess each of these assumptions in turn. The
response criterion given that the distance between the model assumptions are: (i) recollection is a threshold
means of the two normal distributions is d' (see process; (ii) familiarity is a signal-detection process; (iii)
Macmillan & Creelman 1991). The model requires two recollection supports relatively high con¢dence recognition
free memory parameters to generate an ROC; R, which responses; and (iv) these two processes are independent.

Phil. Trans. R. Soc. Lond. B (2001)

1366 A. P.Yonelinas Components of episodic memory

new old item

recognition
0.8

hits ("yes" target)

0.6
"new" "old"
response criterion associative
0.4 recognition

1
0.2
0.8
0
hits ("old" old)

0 0.2 0.4 0.6 0.8 1

0.6
false alarms ("yes" lure)
Figure 4. Recognition memory ROCs for item and associative
0.4 information (Yonelinas 1997, experiment 1).

0.2
ROC is actually a kinked line; when the response
criterion moves to the right of the threshold, the ROC
0
intersects the y-axis and drops. However, as long as the
0 0.2 0.4 0.6 0.8 1
subject places each of their response criteria at ö or to the
false alarms ("old" new) left of ö the threshold, the ROC should be a straight line.
Figure 3. Strength distributions of a high threshold model One way of determining whether recollection re£ects a
and the predicted ROC if performance relies exclusively on threshold process is to look for tests of recognition that
this threshold process. rely primarily on recollection and determine whether
linear ROCs are obtained. Five years ago, it seemed that
this assumption must be incorrect because the previous
Next, I will review the studies that have directly assessed 20 years of recognition memory research had not
these assumptions. produced a single linear ROC. However, these studies
almost always examined only standard old ^ new recogni-
tion judgements, tests in which familiarity could be used
4. DOES RECOLLECTION REFLECT
to discriminate between studied and non-studied items.
A THRESHOLD PROCESS?
In order to test the threshold assumption it is necessary
If recollection is a threshold process, then subjects to ¢nd experimental conditions under which familiarity
either retrieve qualitative information about a previous plays only a limited role in recognition performance.
study event or they fail to. They can, of course, retrieve Such conditions were found in tests of associative recogni-
di¡erent aspects of an event or di¡erent amounts of infor- tion, in which subjects studied pairs of words and were
mation, but if they relax their response criterion below then required to discriminate between previously
the recollective threshold, accurate levels of recollection presented pairs and rearranged pairs (Yonelinas 1997).
will not increase. Figure 3 illustrates the strength distri- Because all the studied and rearranged pairs consisted of
butions of a high threshold model and the predicted ROC familiar items (i.e. they had been studied), familiarity
if performance relies exclusively on this threshold process. was expected to be less useful than in tests of single item
The ROC is generated by moving the response criterion recognition in which the studied items were familiar and
from the right to the left along the strength continuum the non-studied items were novel. If associative recogni-
and accepting the items to the right of the response tion relies primarily on recollection, then the ROCs
criterion as having been studied. The threshold is the should be relatively linear. Figure 4 presents the average
point at which the new item distribution ends (i.e. the ROCs for associative and single item recognition (from
right side of the new item distribution). Note that Yonelinas 1997, experiment 1). Unlike the curvilinear item
threshold models with more than a single threshold may ROC, the associative ROC is relatively linear. The same
be appropriate under some conditions (e.g. Yonelinas results were found in two other experiments in that study
1997), but this single threshold model appears to be su¤- and similar results have since been reported using a
cient to describe recollection in standard recognition variety of di¡erent materials (e.g. Kelley & Wixted 1998;
paradigms. The recollection distributions are discrete or Rottello et al. 2000; Yonelinas et al. 1999b).
square rather than continuously varying, thus, unlike The threshold assumption was further veri¢ed in tests
signal-detection theory, the model generates a linear of source memory (Yonelinas 1999a), in which subjects
ROC. However, note that, strictly speaking, the predicted were required to discriminate between items that had

Phil. Trans. R. Soc. Lond. B (2001)

 Components of episodic memory A. P.Yonelinas 1367

originated from two di¡erent sources (e.g. words spoken 1

by two di¡erent experimenters and words presented in
di¡erent locations). Under test conditions in which the
familiarity of items from the two di¡erent sources were 0.8 controls
not expected to di¡er, the ROCs were relatively linear.
The results of these source and associative memory

hits ("old" old)

studies are important in providing support for the 0.6
assumption that recollection re£ects a threshold process.
Moreover, they indicate that a simple signal-detection
model is not consistent with the ROC data, i.e. signal- 0.4 amnesics
detection theory always predicts curvilinear ROCs, as long
as performance is above chance. Even a two-component
signal-detection model that has separate accuracy and 0.2
variance parameters (i.e. the unequal-variance signal-
detection model) cannot generate linear ROCs.
0
5. DOES FAMILIARITY REFLECT A 0 0.2 0.4 0.6 0.8 1
SIGNAL-DETECTION PROCESS? false alarms ("old" new)

A second critical assumption of the dual-process model is Figure 5. Recognition memory ROCs from amnesics and
that familiarity re£ects an equal-variance signal-detection aged-matched controls (Yonelinas et al. 1998).
process. The most critical aspect of this model is that the
old and new item familiarity distributions are assumed to
have equal variance. There is no a priori reason why this
assumption must be true and there are reasons to think et al. 1998). These results provide support for the claim
that it might be violated. For example, if there is a great that familiarity is well described as an equal-variance
deal of variability in the degree to which studied items signal-detection process, and demonstrate that the model
increase in memory strength due to encoding, then one is useful in understanding the memory performance of
would expect the old item distribution to be associated healthy and memory impaired populations.
with greater variance than the new item distribution. Additional support for the threshold and signal-detection
Alternatively, there may be some upper limit on the assumptions comes from studies using the remember ^
familiarity level that an item can reach, and this could know procedure (Gardiner 1988; Tulving 1985) to
lead the variance of the old item distribution to be less examine the ROCs associated with familiarity and recol-
than that of the new item distribution. lection (e.g. Yonelinas 2001). In the remember ^ know
A way to test the signal-detection assumption directly procedure, subjects are instructed to indicate when a
is to examine recognition performance under conditions recognition judgement is based on recollection (i.e.
in which performance relies exclusively on familiarity. If respond `remember' if you can recollect any qualitative
familiarity re£ects a signal-detection process then the aspect of the study event) and when it is based on famil-
ROC should be curvilinear and symmetrical. One way to iarity in the absence of recollection (i.e. respond `know' if
test this assumption is to examine recognition ROCs in the item is familiar and you know it was studied but you
amnesic patients (e.g. patients with medial temporal lobe cannot recollect anything about the study event). If
damage). Because amnesics are unlikely to recollect subjects are required to make con¢dence judgements and
previous events but are able to make recognition remember ^ know judgements for each test item, then
responses based on assessments of item familiarity remember and know responses can be used to examine
(Huppert & Piercy 1976; Mandler 1980; Mayes 1988), separately the ROCs associated with recollection and
their ROCs should re£ect the contribution of familiarity familiarity.
in the absence of recollection. If the current dual-process Figure 6a shows a recognition ROC derived on the
model is correct, and amnesics are making their recogni- basis of con¢dence judgements (Yonelinas 2001, experi-
tion judgements based on familiarity alone, then their ment 1). In agreement with previous studies, the recogni-
recognition ROCs should be curvilinear and symme- tion ROC is curvilinear and asymmetrical. Figure 6b
trical, in contrast to the asymmetrical functions observed shows the separate ROCs for the remembered items and
in healthy subjects. This prediction was tested by exam- the items accepted on the basis of familiarity. The ¢gure
ining recognition memory for previously studied words in shows that the probability of a remember response
amnesics and healthy control subjects (Yonelinas et al. remained constant as the response criterion was relaxed,
1998). Figure 5 shows that, in contrast to control subjects indicating that relaxing the response criterion below the
who exhibited curved asymmetrical recognition ROCs, recollection threshold did not lead to an increase in
the amnesics' functions were curved and symmetrical. accurate recollection. Familiarity was estimated using the
Note that even when overall recognition performance was independence remember ^ know method of analysis
equated between the two groups by decreasing the study (Yonelinas & Jacoby 1995), i.e. because subjects were
duration of the study items for the control subjects, the instructed to respond `know' whenever an item was
controls still exhibited asymmetrical ROCs in contrast to familiar but not recollected (F(17R)), familiarity was
the amnesics. Similar results have also been observed estimated as the probability of making a `know' response
when recognition memory for faces was tested (Dobbins given that the item was not recollected (F ˆ K/(17R)).

Phil. Trans. R. Soc. Lond. B (2001)

1368 A. P.Yonelinas Components of episodic memory

(a) 6. DOES RECOLLECTION SUPPORT RELATIVELY

1
HIGH CONFIDENCE RESPONSES?
A third assumption underlying the dual-process model
0.8 is that recollection leads to high con¢dence recognition
responses relative to familiarity. This assumption is meant
to capture the notion that when subjects retrieve qualita-
0.6 tive information about a study event they should be con¢-
dent that the event actually occurred. In contrast,
accepting items on the basis of familiarity should be more
0.4 error prone because of the overlapping familiarity distri-
butions of old and new items, thus subjects are expected
to be less sure about familiarity-based responses than
0.2 recollection-based responses. The assumption that recol-
lection supports high con¢dence responses is re£ected in
the manner in which the two components combine in the
0 dual-process model. That is, the model assumes that
0 0.2 0.4 0.6 0.8 1 response criterion should in£uence only familiarity while
hits

(b) recollection should be relatively invariant as the response

1 criterion is relaxed. The assumption that recollection
leads to high con¢dence responses can of course be
violated. For example, if one were to instruct subjects that
0.8 they would be ¢ned $1000 every time they false alarmed
to a new item, they would probably adopt such a strict
familiarity response criterion that they would respond `no' to all the
0.6 familiar items and all the recollected items. The critical
question, however, is whether recollection, in general,
leads to higher con¢dence responses than does familiarity.
0.4 The results in the remember ^ know study described
earlier suggest that recollection does lead to high con¢-
recollection dence responses, i.e. remembered items did not increase
0.2 appreciably as the response criterion was varied, indi-
cating that recollection did not contribute to the lower
con¢dence responses. However, to assess this assumption
0 more directly it is useful to examine the distribution of
0 0.2 0.4 0.6 0.8 1 recognition con¢dence responses for remember and know
false alarms responses. Figure 7a presents the proportion of remember
and know responses to studied items for each level of
Figure 6. (a) Recognition memory ROC, and (b) estimates of
recognition memory con¢dence. It shows that most of
recollection and familiarity derived from remember^know
responses (Yonelinas 2001, experiment 1). the remembered items (94%) led to the highest con¢-
dence recognition responses. In contrast, the know
responses were distributed across the range of response
The ¢gure shows that familiarity increased gradually and con¢dence categories. Note, however, that familiarity
formed a curved and symmetrical function, as expected if based responses were in many cases associated with high
it re£ected an equal-variance signal-detection process. con¢dence responses, and thus con¢dence in itself cannot
Similar results have also been reported using the process be used as an index of recollection and familiarity (for
dissociation procedure (Yonelinas 1994), i.e. when recol- similar arguments, see Gardiner & Java 1990; Rajaram
lection is estimated as the ability to determine list 1993).
membership, recollection is found to remain relatively Similar results were observed in a study in which
constant as the recognition response criterion is relaxed, subjects made recognition con¢dence judgements and
whereas familiarity estimates increase gradually and form source memory judgements (i.e.`was it in list 1 or list 2?')
symmetrical ROCs. for each test item (Yonelinas 2001). In this study, recollec-
The important point of these studies is that they indi- tion was measured, not on the basis of subjective reports
cate that the asymmetrical ROCs that are observed in of remembering but on the basis of accurate source
recognition memory tests arise because both recollection memory, i.e. if the subject can accurately determine when
and familiarity contribute to performance. When the or where the item was studied, it can be assumed that the
recollection-based responses are separated from the item was recollected. In this way, accurate source
familiarity-based responses, familiarity is found to behave memory was used as an index of recollection. Figure 7b
like a classical signal-detection process (i.e. as the response shows the source accuracy associated with each level of
criterion relaxes, familiarity increases and produces a recognition con¢dence. It indicates that items receiving
symmetrical curved ROC), whereas recollection behaves recognition con¢dence scores of less than six were asso-
as a threshold process (i.e. recollection remains relatively ciated with source memory performance that was close to
constant across changes in response criteria). chance (i.e. 50%), whereas items that were recognized

Phil. Trans. R. Soc. Lond. B (2001)

 Components of episodic memory A. P.Yonelinas 1369

(a) familiarity (e.g. Yonelinas 1994; Yonelinas & Jacoby 1995).

0.40 It is important, however, to realize that these latter ¢nd-
0.35 ings do not indicate that the two processes are dependent,
remember and know

remember
0.30 only that some variables play important roles in both
know processes.
responses

0.25
0.20 Other evidence for the independence of recollection
0.15 and familiarity comes from electrophysiological studies of
0.10 recognition. For example, remember and know responses
are correlated with independent event related potentials
0.05
(ERPs). In a study of recognition memory for words,
0
knowing responses were found to be associated with an
(b) early temporo-parietal positivity in the N400 range and
0.9 a late fronto-central negativity (DÏzel et al. 1997). In
0.8 contrast, remembering responses were associated with a
0.7 widespread late bifrontal and a left parieto-temporal posi-
source accuracy

0.6 tivity. Similar recollection and familiarity ERPs have

0.5 been observed in a study in which recollection was
0.4 measured as the ability to recollect the plurality of the
0.3 studied items (Curran 2000), i.e. when subjects were able
0.2 to determine accurately whether the word was studied in
0.1 a singular or plural form, a late temporo-parietal posi-
0 tivity was observed. In contrast, familiar items compared
1 2 3 4 5 6
with new items, regardless of plurality, led to a early posi-
recognition confidence
tivity in the N400 range. Although the ERP results do
not clearly indicate the brain regions supporting recollec-
Figure 7. (a) The proportion of remember and know
responses to studied items for each level of recognition tion and familiarity, the fact that the recollection and
memory con¢dence (Yonelinas 2001, experiment 1). familiarity ERPs are temporally distinct and exhibit
(b) The proportion of studied items leading to correct source distinct scalp topographies suggest that recollection and
judgements for each level of recognition memory con¢dence familiarity rely on partially independent neural generators.
(Yonelinas 1999, experiment 3). Three recent studies using functional magnetic
resonance imaging (fMRI) have provided further evi-
dence that recollection and familiarity involve partially
with the highest level of con¢dence were associated with independent brain regions. For example, we examined the
highly accurate source memory judgements. temporal lobe regions contributing to recognition
These studies indicate that whether recollection is memory for line drawings of objects (Yonelinas et al.
measured as `remember' responses or as the ability to 2001). We found bilateral hippocampal and parahippo-
determine source accurately, items that are recollected campal activation under conditions in which subjects were
are associated with high levels of recognition con¢dence, retrieving associative information accurately about study
whereas familiarity based responses are associated with a items (i.e. the colour it appeared in during the study
wide range of con¢dence responses. phase), relative to conditions under which they were
making accurate item recognition memory judgements
(see Figure 8a). In contrast, item recognition for previously
7. ARE RECOLLECTION AND
studied drawings compared with new drawings was not
FAMILIARITY INDEPENDENT?
associated with hippocampal or parahippocampal activa-
A fourth critical assumption of the dual-process model tion but rather was associated with activation in the left
is that recollection and familiarity are independent inferior temporo-occipital regions (see Figure 8b). Thus,
retrieval processes. This assumption is supported by hippocampal and parahippocampal regions were
numerous behavioural studies indicating that recollection involved in the associative test, in which recollection was
and familiarity are functionally dissociable (for a review required, but were not involved in the old item recogni-
see Jacoby et al. 1997). For example, several variables such tion test, in which familiarity was su¤cient to discrimi-
as amnesia, aging, response dead-lining, dividing atten- nate between studied and non-studied items. In an fMRI
tion and list length have been found to have disproportion- study examining recognition memory for words, the left
ately large e¡ects on recollection compared with hippocampus was associated with greater activation for
familiarity (e.g. Jacoby 1991; Jennings & Jacoby 1993; `remembered' old words compared with correctly rejected
Toth 1996; Yonelinas 2001; Yonelinas & Jacoby 1994; new words, whereas the old words that elicited `know'
Yonelinas et al. 1998). In contrast, variables such as responses did not lead to hippocampal activation relative
response bias, massed priming, and study ^ test lag have to new items (Henson et al. 1999). Remembering and
disproportionately large e¡ects on familiarity (e.g. knowing were also found to involve the frontal and
Rajaram 1993; Yonelinas 1994; Yonelinas & Levy 2001). It parietal cortices di¡erentially in that study, i.e. the left
is of course possible to ¢nd manipulations that have parietal cortex showed greater activation for `remember'
similar e¡ects on both processes. For example, increasing responses while the right lateral and medial frontal cortex
study duration and varying the size of items between showed greater activation for `know' responses. In
study and test appear to in£uence both recollection and another study examining recognition memory for words,

Phil. Trans. R. Soc. Lond. B (2001)

1370 A. P.Yonelinas Components of episodic memory

equations described earlier to observed ROCs, the model

(a) x = −24 can be used to derive estimates of recollection and famil-
P iarity. The method is similar to conducting a linear
H regression in which one ¢ts a line to the observed data
points in order to derive estimates of slope and intercept,
but in this case the function is nonlinear and the estimates
are of recollection and familiarity (for a detailed descrip-
tion of several ¢tting methods, see Yonelinas 1999a). This
method has been used to examine the e¡ects of di¡erent
experimental manipulations on recollection and famil-
iarity (e.g. Yonelinas 2001) and to determine the fate of
(b)
x = −54 O these two processes in di¡erent patient populations (e.g.
Yonelinas et al. 1998).
However, one concern that arises when modelling
T recognition ROCs in this way is that the parameter esti-
mates that are produced may only re£ect a convenient
mathematical description of the ROC data and they may
not capture any real psychological processes. Although it
is important that models provide accurate quantitative
Figure 8. Regions of activation overlaid on the average accounts of existing data, it is equally important that the
normalized T1 structural images. (a) Bilateral hippocampal model's underlying processes are psychologically valid.
[H] and parahippocampal [P] regions were more active The validity of these processes can be assessed by asking
during associative recognition than old item recognition (the whether they correspond with other behaviour measures
left hemisphere activations are shown here). (b) Regions in that are expected to index recollection and familiarity.
the left middle occipital gyrus [O] and left middle temporal For example, recollection should correspond with the
gyrus [T] were more active for old compared with new item ability to determine where or when an item was
recognition. presented, and the estimates derived from the ROC
analysis should thus parallel those derived from the
correct `remember' responses were associated with process dissociation procedure in which recollection is
increases in hippocampal activation while `know' measured as the ability to determine the study source. A
responses were not related to increases in hippocampal related approach is to ask whether these processes have
activation (Eldridge et al. 2000). any phenomenological validity. That is, do they corre-
These neuroimaging studies indicate that the hippo- spond to processes that are available to introspective
campus is involved in recollection, and they suggest that conscious experience. Thus, one can ask whether the esti-
this region is less important for familiarity. The regions mates derived from the ROC analysis correspond to those
found to be involved in familiarity, however, were less derived from the remember ^ know procedure.
consistent across experiments and future studies are To assess these questions one can examine studies that
necessary in order to clearly delineate the anatomical used the ROC procedure, and either the remember ^ know
substrates of this process. Nonetheless, the fMRI results procedure or the process dissociation procedure, and plot
are clear in showing that the two processes do not rely on the estimates derived from the ROC procedure against
identical brain regions, and thus indicate that recollection those derived from the other procedures. If the ROC
and familiarity re£ect distinct memory retrieval method produces estimates of recollection and familiarity
processes. that converge with those derived from the other methods,
Taken together, the behavioural, ERP and fMRI it would indicate that the method is accurately character-
results are consistent with the assumption that recollec- izing the processes underlying recognition memory.
tion and familiarity re£ect independent memory retrieval Figure 9 presents estimates for recollection (a) and famil-
processes. However, it seems quite likely that there may iarity (b) derived from 20 di¡erent experimental condi-
be conditions under which these processes interact and tions from four published studies (Yonelinas 1994;
future studies that aim to examine such interactions will Yonelinas 2001; Yonelinas et al. 1996; Yonelinas & Jacoby
be extremely useful in developing future models of 1995). The bottom axis on each graph represents the esti-
episodic memory. Still, in light of the evidence supporting mates derived from the ROC analysis. The vertical axis
the independence assumption, and the dual-process represents the estimates from the remember ^ know proce-
model's success at accounting for the existing recognition dure (top panels) and process dissociation procedure
memory data, it appears that in general the two processes (bottom panels). The top two ¢gures indicate that the
operate in an independent manner. ROC and remember ^ know procedures produce estimates
that are almost identical (i.e. the points fall along the diag-
onal). The bottom ¢gures show that the estimates from the
8. THE PHENOMENOLOGICAL AND BEHAVIOURAL
ROC and process dissociation procedure are also quite
VALIDITY OF RECOLLECTION AND FAMILIARITY
close. Thus, the estimates from the ROC method
The dual-process signal-detection model can be used in converge with estimates derived using the process disso-
conjunction with recognition memory con¢dence results ciation and remember ^ know procedures.
to derive quantitative estimates for the contribution of Although the three methods produce estimates that are
recollection and familiarity. That is, by ¢tting the model quite close, it is useful to consider the conditions under

Phil. Trans. R. Soc. Lond. B (2001)

 Components of episodic memory A. P.Yonelinas 1371

(a) recollection (b) familiarity

0.7 2
0.6
0.5 1.5

R-K estimates
R-K estimates
0.4
1
0.3
0.2 0.5
0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.5 1 1.5 2

0.7 2
0.6
1.5
PDP estimates

PDP estimates
0.5
0.4
1
0.3
0.2 0.5
0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.5 1 1.5 2
ROC estimates ROC estimates

Figure 9. Comparisons of (a) recollection and (b) familiarity estimates derived from ROC, remember^know and process-
dissociation procedures (PDP) (Yonelinas 1994; Yonelinas 2001; Yonelinas et al. 1996; Yonelinas & Jacoby 1995).

which the procedures do di¡er. For example, the two The utility of a memory theory is determined in part
points in the bottom left panel of ¢gure 9 that fall the by its ability to reveal hidden order in otherwise complex
farthest below the diagonal re£ect estimates of recollec- datasets. The dual-process model succeeds in doing this
tion derived from a study that used the ROC and process by revealing the direct relationship between ROC,
dissociation procedures (Yonelinas & Jacoby 1995). The process dissociation and remember ^ know paradigms. It
discrepancy in that study is probably due to the fact that shows that there are two processes, recollection and famil-
the test conditions in the two tasks were not identical. iarity, that underlie episodic recognition memory, and
Most important was that the list discrimination required that these two processes are responsible for the complex
in the process dissociation task was particularly di¤cult patterns of results that we see in recognition ROCs, tests
because the encoding conditions in the two study lists of associative or source recognition, and in subjective
were similar. Because the process dissociation procedure reports of remembering and knowing.
used in that study measured recollection as the ability to
discriminate between these two similar lists, the sub-
9. MODEL LIMITATIONS
sequent estimates of recollection were quite low. In
contrast, in the ROC con¢dence procedure, because list Despite the model's successes, it is quite clear that it is
discrimination was not required, any information that the insu¤cient. Its most obvious limitation is that it is too
subject remembered could be used as a basis for recollec- simple. That recognition memory performance could
tion and thus the estimates of recollection were higher. be accounted for with two or three free memory para-
Thus, discrepancies can arise between these di¡erent meters is extremely unlikely, and there will undoubtedly
methods of measuring recollection and familiarity when be cases in which additional processes and alternative
the processes are measured under di¡erent conditions. assumptions will be required. Preliminary evidence that
However, there is generally good agreement across the the model may be too simple comes from studies indi-
procedures when the conditions are held constant. cating that it is sometimes found to deviate slightly from
The convergence of the results from the three di¡erent the observed recognition ROC data (e.g. Ratcli¡ et al.
methods indicates that it is not necessary to plot an entire 1995; Glanzer et al. 1999; Yonelinas 1999b, 1994). The
ROC in order to determine its shape. Rather, asking a deviation takes the form that the ROCs are sometimes
subject to make remember ^ know judgements or source slightly more curved than the dual-process model
memory judgements appears to provide the same infor- predicts. Note that the same problem arises for the
mation. Conversely, it does not appear to be necessary to unequal-variance signal-detection model as well.
ask subjects to report on the subjective experiences of However, the observed deviations from the dual-process
recollection and familiarity in order to determine the model are quite subtle: typically the observed points and
likelihood that subjects will have these conscious experi- the points predicted by the model deviate by about 0.01
ences. Rather the ROC or process dissociation procedures or 0.02. Nonetheless, these deviations may be important
can be used in conjunction with the dual-process model to and may indicate that a noise parameter or guessing
predict the occurrence of these conscious states. process is needed (for a discussion of possible explanations

Phil. Trans. R. Soc. Lond. B (2001)

1372 A. P.Yonelinas Components of episodic memory

for these deviations, see Ratcli¡ et al. 1994; Yonelinas requires careful consideration of the task demands of each
1999b). memory test.
Related issues may arise in tests of associative and Probably the most critical limitation of the current
source recognition. Although associative and source model is that it does not specify how memories are
recognition ROCs can be linear, there are cases in which represented or how these processes are neurally instant-
these ROCs are noticeably curved. One possible explana- iated. Although it is broadly consistent with some neuro-
tion for these ¢ndings is that a signal-detection process is anatomical models that postulate that recollection and
contributing to these judgements. For example, there may familiarity processes are supported by distinct temporal
be conditions under which recollection behaves in a more lobe regions (e.g. Aggleton & Brown 1999; Eichenbaum
continuous manner, or in which familiarity supports et al. 1994; O'Reilly et al. 1997), the neural substrates of
source or associative memory judgements. Preliminary recollection and familiarity are not yet known. Careful
studies have already begun to investigate these issues. For consideration of neuroanatomy and neurophysiology will
example, one obvious case in which familiarity can be essential in future developments of any episodic
support source memory judgements is when one list of memory model.
items is presented much earlier than another. In this case, The approach taken here is to begin with a simple
subjects may accept the more familiar items as having quantitative model and to carefully test the model's
originated in the more recent list. As expected, source assumptions. The idea is that once the basic assumptions
memory ROCs under these conditions tend to be curved are veri¢ed and the boundary conditions under which
(e.g. see Yonelinas 1999a). these assumption hold are determined, additional
Familiarity can also contribute to associative memory assumptions can be added or the existing assumptions
judgements under some conditions. For example, if asso- can be modi¢ed. This approach di¡ers from the two
ciative information is `unitized' during encoding then approaches that have dominated recent cognitive research
familiarity may support associative judgements. That is, if in memory. One approach has been to propose complex
the subject treats two aspects of a study event as a whole quantitative models that require numerous assumptions
unit, or gestalt, then that whole unit, as well as its constit- about how items are represented and how the retrieval
uent parts, may become familiar. For example, when mechanisms work (e.g. the global memory models). The
subjects are required to discriminate between repeated advantage of this approach is that the models make quan-
faces and rearranged faces (e.g. the internal features of a titative predictions that can be directly tested. The
studied face, such as the eyes, nose and mouth, are paired complexity of these models, however, has in general
with external features, such as the hair, ears and chin, of precluded the possibility of testing their individual
another studied face), because each face is treated as a assumptions. Moreover, these models have in general
holistic unit (for a review, see Searcy & Bartlett 1996) a focused on behavioural results rather than on neurobiolo-
repeated face should be more familiar than a mixed face. gical ¢ndings and thus they tend to say very little about
Thus subjects may make use of familiarity to make the the data coming from neuropsychological and neuroima-
associative memory judgement, leading to curvilinear ging studies. An alternative approach that is dominant in
ROCs. As expected, the associative ROCs under these cognitive neuroscience studies of memory is to propose
conditions are found to be curvilinear (Yonelinas et al. general theoretical frameworks that are designed to
1999). Note, however, that when the faces are studied and capture the important distinctions seen in the neuropsy-
tested upside down, each face is no longer treated as chological and neuroimaging literatures. For example,
holistic unit and the resulting associative ROCs are theories proposing distinctions between episodic and
linear. semantic memory (Tulving 1983), or between declarative
Familiarity can also support associative memory for and procedural memory (Squire 1987), have been useful
word pairs as long as the two words are treated as a in guiding research and relating human and non-human
single unit. For example, when two words form a studies of memory, but because they are generally not
compound word (e.g. sea-food, o¡-shore, ice-cube), quantitative models their predictive power has been
repeated word pairs (e.g. ice-cube) are more familiar limited.
than new word pairings (e.g. sea-shore) and the resulting The dual-process model re£ects a theory that lies
ROCs are also curvilinear ( J. Quamme & A. P. somewhere between these two dominant approaches and
Yonelinas, unpublished data). The same e¡ects are also I would like to believe that it builds on the strengths asso-
observed when the word pairs do not form pre-existing ciated with each approach. The model is a gross over-
compound words (e.g. sea-cube), as long as the subjects simpli¢cation of the processes that subjects bring to bear
encode the pair as a coherent whole (e.g. at study they in episodic memory tests. However, the model does
are instructed to generate a de¢nition for the novel provide a very simple and powerful tool for under-
compound word sea-cube). standing memory performance in a variety of recognition
These studies indicate that familiarity may be more memory paradigms, and it does point to a fundamental
£exible than was originally thought and that there may distinction between two di¡erent types of recognition
be conditions under which it can support recognition retrieval processes.
judgements previously thought to require recollection.
These results are important in reminding us that memory
10. CONCLUSIONS
tests should not be treated as direct measures of under-
lying memory processes (for similar arguments, see Dissociations observed in recognition memory perfor-
Jacoby 1991) and that determining the contribution of mance indicate that there are at least two components of
recollection and familiarity to memory performance episodic memory. A dual-process model that assumes that

Phil. Trans. R. Soc. Lond. B (2001)

 Components of episodic memory A. P.Yonelinas 1373

subjects can make recognition responses on the basis of Jacoby, L. L. 1991 A process dissociation framework: separating
independent recollection and familiarity processes is automatic from intentional uses of memory. J. Mem. Lang. 30,
found to be consistent with behavioural, neuropsycho- 513^541.
logical and neuroimaging studies of recognition memory. Jacoby, L. L. & Dallas, M. 1981 On the relationship between
autobiographical memory and perceptual learning. J. Exp.
The results indicate that recollection is well described as
Psychol. Gen. 110, 306^340.
a threshold process, whereby qualitative information Jacoby, L. L., Yonelinas, A. P. & Jennings, J. M. 1997 The
about previous study events is retrieved, whereas famil- relationship between conscious and unconscious (automatic)
iarity is well described as a classical signal-detection in£uences: a declaration of independence. In Scienti¢c
process, whereby familiar items are accepted as having approaches to the question of consciousness (ed. J. Cohen &
been studied. J. W. Schooler), pp. 13^47. Hillsdale, NJ: Earlbaum.
Jennings, J. M. & Jacoby, L. L. 1993 Automatic versus inten-
The work was supported by grant MH59352 from the National tional uses of memory: aging, attention, and control. Psychol.
Institute of Mental Health Bethesda, MD, USA. Aging 8, 283^293.
Kelley, R. & Wixted, J. T. 2001 On the nature of associative
information in recognition memory. J. Exp. Psychol. Learn.
REFERENCES Mem. Cogn. 27, 701^722.
Aggleton, J. P. & Brown, M. B. 1999 Episodic memory, amnesia Macmillan, N. A. & Creelman, C. D. 1991 Detection theory: a
and the hippocampal ^anterior thalamic axis. Behav. Brain Sci. user's guide. Cambridge University Press.
22, 425^490. Mandler, G. 1980 Recognizing: the judgment of previous occur-
Atkinson, R. C. & Juola, J. F. 1974 Search and decision rence. Psychol. Rev. 87, 252^271.
processes in recognition memory. In Contemporary developments Mayes, A. R. 1988 Human organic memory disorders. Cambridge
in mathematical psychology. Vol. 1. Learning, memory & thinking (ed. University Press.
D. H. Krantz, R. C. Atkinson, R. D. Luce & P. Suppes). San Murdock, B. B. 1982 A theory of storage and retrieval of item
Francisco: Freeman. and associative information. Psychol. Rev. 89, 609^666.
Clark, S. E. & Gronlund, S. D. 1996 Global matching models of Murdock, B. B. & Dufty, P. O. 1972 Strength theory and recog-
recognition memory: how the models match the data. Psychon. nition memory. J. Exp. Psychol. Gen. 94, 284^290.
Bull. Rev. 5, 37^60. O'Reilly, R. C., Norman, K. A. & McClelland, J. L. 1997 A
Curran, T. 2000 Brain potentials of recollection and familiarity. hippocampal model of recognition memory. Proceedings of the
Mem. Cognit. 28, 923^938. Neural Information Processing Systems Meeting, CS7, p. 73.
Dobbins, I., Yonelinas, A. P., Kroll, N. E. A, Soltani, M. & Rajaram, S. 1993 Remembering and knowing: two means of
Knight, R. T. 1998 Recognition memory in amnesia and access to the personal past. Mem. Cogn. 21, 89^102.
healthy subjects. Paper presented at the Annual Meeting of the Ratcli¡, R., Sheu, C. F. & Gronlund, S. D. 1992 Testing global
Cognitive Neuroscience Society, San Francisco. memory models using ROC curves. Psychol. Rev. 3, 518^535.
Donaldson, W. & Murdock, B. B. 1968 Criterion change in Ratcli¡, R., McKoon, G. & Tindall, M. 1994 Empirical gener-
continuous recognition memory. J. Exp. Psychol. 76, 325^330. ality of data from recognition memory receiver-operating
DÏzel, E., Yonelinas, A. P., Mangun, G. R., Heinze, H. & characteristic functions and implications for the global
Tulving, E. 1997 Event-related brain potential correlates of memory models. J. Exp. Psychol. Learn. Mem. Cogn. 20, 763^785.
two states of conscious awareness in memory. Proc. Natl Acad. Ratcli¡, R., Van Zandt, T. & McKoon, G. 1995 Process dissocia-
Sci. USA 94, 5973^5978. tion, single process theories, and recognition memory. J. Exp.
Eichenbaum, H., Otto, T. & Cohen, N. 1994 Two functional Psychol. Gen. 124, 352^374.
components of the hippocampal memory system. Behav. Brain Rotello, C. M., Macmillan, N. A. & Van Tassel, G. 2000 Recall-
Sci. 17, 449^518. to-reject in recognition: evidence from ROC curves. J. Mem.
Eldridge, L. L., Knowlton, B. J. Furmanski, C. S., Bookheimer, Lang. 43, 68^88.
S. Y. & Engel, S. A. 2000 Remembering episodes: a selective Searcy, J. H. & Bartlett, J. C. 1996 Inversion and processing of
role of the hippocampus during retrieval. Nature Neurosci. 3, component and spatial-relation information in faces. J. Exp.
1149^1152. Psychol. Hum. Percept. Perform. 22, 904^915.
Gardiner, J. M. 1988 Functional aspects of recollective experi- Squire, L. R. 1987 Memory and brain. New York: Oxford
ence. Mem. Cognit. 16, 309^313. University Press.
Gardiner, J. M. & Java, R. I. 1990 Recollective experience in Toth, J. P. 1996 Conceptual automaticity in recognition
word and nonword recognition. Mem. Cognit. 18, 23^30. memory: levels-of-processing e¡ects on familiarity. Can. J.
Glanzer, M., Kim, K., Hilford, A. & Adams, J. K. 1999 Slope Exp. Psychol. 50, 123^138.
of the receiver operating characteristic in recognition Tulving, E. 1983 Elements of episodic memory. Oxford University
memory. J. Exp. Psychol. Learn. Mem. Cogn. 25, 500^513. Press.
Henson, R. N. A., Rugg, M. D., Shallice, T., Josephs, O. & Tulving, E. 1985 Memory and consciousness. Can. Psychol. 26,
Dolan, R. J. 1999 Recollection and familiarity in recognition 1^12.
memory: an event-related functional magnetic resonance Yonelinas, A. P. 1994 Receiver-operating characteristics in
imaging study. J. Neurosci. 19, 3962^3972. recognition memory: evidence for a dual-process model. J.
Hintzman, D. L. 1986 `Schema abstraction' in a multiple trace Exp. Psychol. Learn. Mem. Cogn. 20, 1341^1354.
memory model. Psychol. Rev. 94, 341^358. Yonelinas, A. P. 1997 Recognition memory ROCs for item and
Hockley, W. E. 1991 Recognition memory for item and associa- associative information: evidence for a dual-process signal-
tive information: a comparison of forgetting rates. In Relating detection model. Mem. Cogn. 25, 747^763.
theory and data: essays on human memory in honor of Bennet B. Yonelinas, A. P. 1999a The contribution of recollection and
Murdock (ed. W. E. Hockley & S. Lewandowsky), pp. 227^248. familiarity to recognition and source memory judgments: a
Hillsdale, NJ: Erlbaum. formal dual-process model and an ROC analysis. J. Exp.
Huppert, F. & Piercy, M. 1976 Recognition memory in amnesic Psychol. Learn. Mem. Cogn. 25, 1415^1434.
patients: e¡ects of temporal context and familiarity of mate- Yonelinas, A. P. 1999b Recognition memory ROCs and the
rial. Cortex 12, 3^20. dual-process signal detection model: comment on Glanzer,

Phil. Trans. R. Soc. Lond. B (2001)

1374 A. P.Yonelinas Components of episodic memory

Kim, Hilford and Adams (1999). J. Exp. Psychol. Learn. Mem. H. S. & King, L. 1996 Signal-detection, threshold, and dual-
Cogn. 25, 514^521. process models of recognition memory: ROCs and conscious
Yonelinas, A. P. 2001 The three Cs of recognition memory: recollection. Conscious. Cogn. 5, 418^441.
consciousness, control and con¢dence. J. Exp. Psychol. Gen. (In Yonelinas, A. P., Kroll, N. E. A., Dobbins, I. G., Lazzara, M. &
the press.) Knight, R. T. 1998 Recollection and familiarity de¢cits in
Yonelinas, A. P. & Jacoby, L. L. 1994 Dissociation of processes amnesia: convergence of remember/know, process dissocia-
in recognition memory: e¡ects of interference and of test tion, and ROC data. Neuropsychology 12, 323^339.
speed. Can. J. Exp. Psychol. 48, 516^534. Yonelinas, A. P., Kroll, N. E. A., Dobbins, I. G., Soltani, M. 1999
Yonelinas, A. P. & Jacoby, L. L. 1995 The relation between Recognition memory for faces: when familiarity supports asso-
remembering and knowing as bases for recognition: e¡ects of ciative recognition judgments. Psychonom. Bull. Rev. 6, 654^661.
size congruency. J. Mem. Lang. 34, 622^643. Yonelinas, A. P., Hop¢nger, J. B., Buonocore, M. H., Kroll.,
Yonelinas, A. P. & Levy, B. 2001 Dissociating familiarity from N. E. A., & Baynes, K. 2001 Hippocampal, parahippocampal
recollection in human recognition memory. (Submitted.) and occipital-temporal contributions to associative and item
Yonelinas, A. P., Dobbins, I., Szymanski, M. D., Dhaliwal, recognition memory: An fMRI study. Neuroreport 12, 2001.

Phil. Trans. R. Soc. Lond. B (2001)