Awake replay of remote experiences inthe hippocampus

Mattias P Karlsson & Loren M Frank

Hippocampal replay is thought to be essential for the consolidation of event memories in hippocampal-neocortical networks.

Replay is present during both sleep and waking behavior, but although sleep replay involves the reactivation of stored

representations in the absence of specific sensory inputs, awake replay is thought to depend on sensory input from the current

environment. Here, we show that stored representations are reactivated during both waking and sleep replay. We found frequent

awake replay of sequences of rat hippocampal place cells from a previous experience. This spatially remote replay was as common

as local replay of the current environment and was more robust when the rat had recently been in motion than during extended

periods of quiescence. Our results indicate that the hippocampus consistently replays past experiences during brief pauses in

waking behavior, suggesting a role for waking replay in memory consolidation and retrieval.

The hippocampus is essential for the formation of long-termmemories for events1,2. This process is thought to involve rapidencoding in highly plastic hippocampal circuits followed by aconsolidation process in which hippocampal representations are‘reactivated’, allowing these patterns to be engrained in less-plastichippocampal-neocortical circuits3,4. Reactivation is hypothesized todepend on hippocampal sharp wave–ripple (SWR) events5–8

because SWRs propagate from the hippocampus to adjacent corticalregions9,10 and include firing patterns that are associated withprevious6,11–16 and current experiences17–19.

Reactivation has been observed in ensembles of simultaneouslyrecorded hippocampal place cells. These spatial reactivation eventshave been separated into two categories: events that occur outside of anenvironment (following the experience) and events that occur in theenvironment (during the experience). Studies of reactivation outside ofthe reactivated environment have focused largely on sleep and extendedperiods of awake immobility where reactivation can occur in theabsence of the original sensory inputs. During these periods, hippo-campal place cells that fire together during exploration tend to firetogether afterwards during SWRs6,11–13. Complementary studies haveshown that entire sequences can be replayed at high speeds14–16,19.

In contrast, reactivation in the explored environment has beenassociated with current sensory inputs17–19. Recent studies haveobserved awake replay events beginning with the activation of cellswhose place fields were close to the animal and progressing to cells withplace fields that were farther away. These observations led to thehypothesis that awake replay, unlike replay in sleep or sleep-like states,is a result of sequential activation of sensory-driven place fields. Onemodel suggests that the progressive increase in hippocampal depolar-ization during a SWR9,10 causes cells with place fields close to the

animal (and thus with membrane potentials close to threshold) to firefirst. Cells with place fields farther away start from a less-depolarizedpotential and would therefore require more input and a longer time tobecome active17–19.

But is awake replay always dependent on sensory inputs? If awakeand sleep replay instead both activate stored representations of pastexperiences, we would expect that animals could replay experiences ofone place while being awake in a different place. We therefore examinedreplay of hippocampal place-cell sequences from the CA3 and CA1regions of the hippocampus during waking experience across multipleenvironments. We found robust remote replay of past experiencesduring waking behavior, consistent with a role for awake replay inmemory retrieval and consolidation.

RESULTS

We recorded ensembles of principle neurons from hippocampal areasCA3 and CA1 while rats were sequentially exposed to two physicallydifferent W-shaped environments (run sessions in environments 1and 2) and during intervening sessions in a high-walled rest box(Fig. 1)20. Each rat was exposed to environment 2 for two run sessionsper day for either 3 (rat 1) or 6 (rats 2 and 3) consecutive days beforethe first exposure to environment 1, so environment 1 was always morenovel than environment 2. The environments were oriented at90 degrees with respect to one another and were separated by a highbarrier so that the rat had access to largely distinct sets of visual cuesfrom each environment (Supplementary Fig. 1 online). This layouthelped to ensure that the two environments were associated withdistinct hippocampal representations (see below). Each environmenthad one reward site at the endpoint of each arm and the rats wererewarded for performing a continuous alternation task20–22. Data

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Received 25 February; accepted 28 April; published online 14 June 2009; doi:10.1038/nn.2344

W.M. Keck Center for Integrative Neuroscience and Department of Physiology, University of California, San Francisco, California, USA. Correspondence should be addressedto L.M.F. (loren@phy.ucsf.edu).

NATURE NEUROSCIENCE ADVANCE ONLINE PUBLICATION 1

ART ICLES

presented here were recorded on the day of the first exposure toenvironment 1 and the days thereafter (Supplementary Fig. 2 online).

We restricted our analyses to putative excitatory neurons with clearplace-specific firing and stable clusters. To avoid confusing replayevents and sequential firing during movement-related phase preces-sion23, we examined SWRs that occurred when rats were moving lessthan 2 cm s�1. Individual cells had very similar patterns of spatialactivity across the two exposures to environment 1 but generally hadvery different patterns in environments 1 and 2 (Fig. 1b and Supple-mentary Fig. 3 online). We initially combined CA3 and CA1 cells into asingle population to maximize the number of neurons that were activeduring each SWR.

Awake replay in environment 1

Place specificity led to sequences of neural activity as the rat movedthrough the environment (Fig. 2a,b). Replay of these sequences couldonly be detected if a sufficient number of cells with place fields wereactivated, so we defined candidate replay events as SWRs activating fiveor more cells with place fields in environment 1. We divided candidateevents into 15-ms bins and used a Bayesian decoder24,25 with a uniformprior to translate the ensemble spiking into probability distributionsover positions in environment 1. We then determined the likelihoodthat the ordered firing seen during a candidate event corresponded to acoherent spatial sequence. Sequences of ordered positions that wereunlikely to occur by chance (P o 0.05) were considered to representstatistically significant replay (see Online Methods).

We found robust and frequent awake replay of environment 1 duringboth the first (Fig. 2c–f) and second (Supplementary Table 1 online)exposure on each day. We ordered place cells by the distance of eachcell’s place field peak from the endpoint of the center arm of environ-ment 1 (refs. 17–19) (Fig. 2d). Thus, individual replay events werevisible as diagonal sweeps of spike trains. On rare occasions, the trajec-tory corresponding to a sweep went from one outer arm to the other, inwhich case the cells were ordered as a function of the distance from theend of an outer arm. Overlapping SWRs were combined across tetrodes,so many events extended beyond the SWR seen on a single tetrode.

In total, 288 of 612 (47%) candidate events during the firstrun session in environment 1 involved statistically significant replay(Po 0.05) of environment 1. We refer to these events as ‘local replay’ todistinguish them from ‘remote replay’, where sequences from environ-ment 1 were replayed while the rat was in a different environment (seebelow). The proportion of significant events was much greater than the5% expected by chance (Z¼ 20.85, Po 10�10) and was similar to thatseen in previous studies17,18. These events occurred throughout the

track and were prevalent at the choice point, where rats coming fromthe food well on the center arm of the maze had to choose an outsidearm to visit (Supplementary Fig. 4 online). Thus, these events may besimilar to the vicarious trial and error activity reported previously26

(see Supplementary Results online). As in previous studies of novelenvironments27 and awake replay17, the majority of the place cellsrecorded were active in both directions of motion (for example,Fig. 2a). We chose not to apply a specific criterion to separateunidirectional and bidirectional cells and thus we did not classifyevents as being either forward or reverse replay17,18.

Awake and quiescent replay of environment 1 in the rest box

We found robust and frequent replay of environment 1 in SWRsoccurring during awake periods in the rest box. Of the four daily restsessions, we first combined data from Rest 2 and Rest 3, as these bothimmediately followed run sessions in environment 1. Behavior in therest box alternated between periods of movement and periods ofextended immobility. To be conservative, we defined awake periodsas times when rats had been immobile no more than 5 s. In SWRsoccurring during these periods, we observed clear sequential replay ofenvironment 1 place fields (Fig. 2g–n). Of the candidate events, 256 outof 580 (44.1%) showed significant replay of environment 1 duringawake periods in the rest box (Z ¼ 15.48, P o 10�10). These awakereplay events occurred consistently across all three rats (rat 1, 5 of 22events; rat 2, 149 of 346 events; rat 3, 102 of 212 events; all 45%, Po0.05). The proportion of significant events was not different than the47% for run sessions in environment 1 (rest versus environment 1, Z¼1.01, P 4 0.1, not significant), and these events represented physicallypossible trajectories on the track (see Supplementary Results).

We then compared the prevalence of replay during these awakeepisodes with replay during more extended periods of immobility inthe rest box. In the absence of electromyography data, we chose to beconservative and labeled these periods as quiescent. We first definedperiods of quiescence as times when rats had been immobile for 5 s orlonger. This threshold was chosen to capture short periods of sleep(Supplementary Fig. 5 online), but probably also included a numberof awake periods.

The commonly accepted idea that memory reactivation occursprimarily during sleep-like states led us to expect that replay activitywould be most robust during quiescence. In actuality, replayof environment 1 during quiescent SWRs was less likely to occur;373 of 1,046 (35.7%) candidate events showed significant replay (awake(44.1%) 4 quiescence, Z ¼ 3.36, P o 0.001; Fig. 3a). We also used amore stringent criterion for quiescence (immobility for more than 60 s)

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Figure 1 Overview of experimental design.

(a) Sequence of sessions in each day. Each day of

recording consisted of two 15-min exposures to

environment 1 (E1) followed by one 15-min

exposure to environment 2 (E2). Each exposure

was flanked by 20-min rest sessions in the rest

box. The total size of each W track was 76 cm2

and the width of the arms was 7 cm. The restbox was 25 � 34 cm. (b) Distinct spatial

representations for environment 1 and

environment 2. Each column shows the spatial

rate maps for one neuron for both environment

1 (top) and environment 2 (bottom). The number

to the upper right of each plot corresponds to the

maximum rate shown for that cell, which was

65% of the neuron’s peak spatial rate. The color bar to the right illustrate the range of colors that are mapped from 0 to the maximum displayed rate. Rates

were rounded to the nearest whole number. The plots show 10 out of 33 simultaneously recorded neurons with environment 1 and/or environment 2 place

fields from rat 3 on day 8.

Restbox

(20 min)

E1(15 min)

Restbox

(20 min)

E1(15 min)

Restbox

(20 min)

E2(15 min)

Restbox

(20 min)

Rest 1 Rest 2 Rest 4Rest 3

E1

E2

a

b Cell 1 2 3 4 5 6 7 8 9 10

Max

0

22

0

19

16

3

0

1

18

0

8

27

0

0

20

5

0

17

16

5

0

2 ADVANCE ONLINE PUBLICATION NATURE NEUROSCIENCE

ART ICLES

and obtained the same results (106 of 310 events, 34.2%, awake 4quiescence, Z ¼ 2.88, P o 0.005).

The observed difference was not a result of a lack of activity fromenvironment 1 place cells during quiescence, as rates were higherduring quiescent as compared to awake SWRs for 217 of 244 cells(Z ¼ 9.34, P o 10�10; Fig. 3b). Given the occurrence of a candidateevent, however, fewer place cells were active in individual quiescentevents than in individual awake events (rank sum, Z¼ 6.65, Po 10�10;Fig. 3c). Thus, although the overall rate of spiking during SWRs waslower in the awake state, neuronal activation tended to be moreconcentrated in select SWRs.

Although the comparison of the number of replay events for thesame population of place cells is meaningful, the absolute number ofreplay events detected on a day will depend on the total number ofneurons with place fields recorded. We therefore developed a comple-mentary pair-wise analysis approach that is much less sensitive to thesesampling issues. We calculated an R2 value that describes the extent towhich the distance between the peaks of two neurons’ place fieldspredicts the timing of their SWR spikes (see Online Methods). We alsoused a modified version of that analysis to visualize pair-wise activity.We computed cross-correlation histograms of spike trains from all pair-wise combinations of simultaneously recorded place cells and plottedthe normalized histograms as a function of the linear distance betweenthe place field peaks28 (see Online Methods). The pair-wise signature ofreplay is an expanding V shape centered at 0-ms latency. The resultingplots for awake and quiescent SWRs suggested that awake SWRs weremore structured (Fig. 3d).

We quantified that impression and foundthat environment 1 place field distances weremore strongly predictive of the time betweenSWR spikes during wakefulness than quies-cence at both the 5- and 60-s immobility

thresholds (awake: R2 ¼ 0.1164, 70,803 spike pairs, 5,442 SWRs;quiescent 5-s threshold: R2 ¼ 0.0693, 99,598 spike pairs, 28,003 SWRs;quiescent 60-s threshold: R2 ¼ 0.0875, 23,908 spike pairs, 15,074 SWRs;awake 4 quiescent P o 10�10). The two quiescent R2 values were alsosignificantly different (Z¼ 4.90, Po 10�7), suggesting that the strengthof replay varies as a function of the length of immobility at a pair-wiselevel. The difference in awake and quiescent replay could not beexplained by clustering errors, cluster instability or a decay in replaystrength as a function of the time since the environment 1 experience(see Supplementary Results and Supplementary Clusters online).

Awake replay of environment 1 in environment 2

We then asked whether environment 1 replay could continue duringperformance of the alternation task in environment 2. Despite thedistinct representations of environment 1 and environment 2, 182 of442 (41.2%) candidate replay events showed statistically significantreplay (P o 0.05) of environment 1 when rats were located inenvironment 2 (Z ¼ 12.67, P o 10�10; Fig. 4a–l and SupplementaryFigs. 6 and 7 online). As expected, there were also many replays ofenvironment 2; 147 of 330 (44.6%) candidate events were significant(Z ¼ 14.45, P o 10�10). Replay in environment 2 was not affected byimmobility time (see Supplementary Results and SupplementaryFigs. 5 and 8 online), suggesting that any brief periods of quiescencein environment 2 had little effect on the overall proportion of replayevents of either environment 1 or environment 2.

We confirmed these results using our pair-wise analysis, whichshowed a clear relationship between environment 1 place field distance

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

020406080100120140160180Decoded distance to center endpoint (cm)

100 ms

200 ms

100 ms

Time

E1

plac

e ce

lls

1

5

10

15

Pro

babi

lity

0

0.2

0.4

k l m n

a

E1trajectory

E1replay

Decoded distance to center endpoint (cm)0 20 40 60 80 100 120 140 160 180Time

E1

plac

e ce

lls

15

10

15

Pro

babi

lity

0

0.2

0.4

b

c d e f

Rest boxtrajectory

E1replay

E1replay

020406080100120140160180Decoded distance to center endpoint (cm)

E1

plac

e ce

lls

Time1

5

10

15

Pro

babi

lity

0

0.1

0.2

g h i jRest boxtrajectory

1

10

2

11 12

3 4

13

5

14

6

15

7

16

8

17

9

18

Cel

l num

ber

Time1

5

10

15

5 sec0

170

Lineardistance (cm

)

Max

0

14

3

10

9

6

13

13

22

16

18

9

18

13

18

13

8

19

15

Figure 2 Spatially remote awake replay in the

rest box. (a) Spike rasters (black) from 18 cells

active in environment 1 along a correct and a

subsequent incorrect trajectory. The same cells

and associated numbers are used in all panels.

The red line shows the rat’s linear distance during

the same period and the W track cartoons below

show the specific locations that the rat traversed.(b) Two-dimensional spatial rate maps for the

environment 1 place cells that were active in a.

(c) The rat’s trajectory during and after a replay

event. The grey dots represent all of the sample

locations, the yellow circle represents the rat’s

location during the SWR and the green dots

represent the rat’s location in the 5 s following

the SWR. Here the rat was still for more than 5 s

before the SWR, but in g and k, red dots represent

the locations during the 5 s before the SWR.

(d) Sequential spiking during the SWR. Bottom,

rasters of all environment 1 place cells that were

activated during the SWR. Top, the filtered local

field potential (LFP) signal from one tetrode. The

color bar shows the colors associated with each of

the 15-ms decoding bins. (e) Decoded locations

for each bin. Each colored line represents the

probability distribution resulting from decoding

the spiking in the associated 15-ms period from d.(f) A cartoon of the replayed trajectory in

environment 1. (g–n) Examples of awake replay of

environment 1 in the rest box. The motion before

and after the events demonstrates that the rat

was awake.

NATURE NEUROSCIENCE ADVANCE ONLINE PUBLICATION 3

ART ICLES

and pair-wise spike timing of environment 1 place cells during SWRs inenvironment 2 (Fig. 4m). We focused on environment 1 place cells thatdid not have place fields in environment 2; these cells were activeprimarily during SWRs. The pair-wise regression analysis for environ-ment 1 cells yielded a higher R2 value in environment 2 than for awakeperiods in the rest box (regression: R2 ¼ 0.1736, 19,850 spike pairs,18,512 SWRs, compared with rest; Z ¼ 8.57, P o 10�10). Finally, wealso applied the pair-wise measure to CA3 and CA1 neurons separatelyand found clear evidence for sequential activity in both regions, as wellas stronger pair-wise correlations in CA3 than in CA1 (see Supple-mentary Results and Supplementary Fig. 9 online).

The prevalence of environment 1 replay events in environment 2 ledus to ask whether environment 1 replay was related to replay ofenvironment 2. Although most significant replay events replayed onlyone track (Fig. 4e,f,k,l), 24 replayed both tracks simultaneously (24 of182 replay events for environment 1 (13.2%) and 24 of 147 events forenvironment 2 (16.3%)). To determine whether environment 1 andenvironment 2 replay occurred together more often than expected bychance, we examined all of the events that were candidates for bothenvironment 1 and environment 2 replay. We found that the P valuesfor environment 1 and environment 2 decoding regressions were notcorrelated (R ¼ 0.07, P 4 0.05, not significant; see Online Methods).We also asked, for each event that was a candidate for both environ-ment 1 and environment 2, whether the P value for the environment 1replay was related to the number of environment 2 cells active, as mightbe expected if coherent environment 1 replay suppressed environment2 activity. The correlation was not significantly different than 0(R ¼ 0.05, P 4 0.05, not significant). Thus, joint replay appears toresult from independent replay of environment 1 and environment 2.

We then asked whether the direction of replay in environment 1 wasrelated to the rat’s location in environment 2. We reasoned that, attimes when a rat was located at one of the three endpoints of the Wtrack, replay initiated at the rat’s location must represent travel in thedirection extending away from that endpoint. Thus, we used thedirection of the decoded replay (toward or away from the center armfood well) to determine whether replay of environment 1 could haveinitiated at a corresponding location in environment 2.

Although most environment 1 replay events that occurred when therat was in environment 1 moved away from the rat’s position17–19 (151of 190 events, 79.5%; more than half: Z ¼ 6.01, P o 10�8), 20.5% did

not (Supplementary Fig. 10 online). Furthermore, environment 1replay events in environment 2 moved away from the rat’s correspond-ing location in environment 1 only about half the time (85 of 158events, 53.8%, not greater than 50%, Z ¼ 0.68, P 4 0.1, notsignificant). Thus, nearly half of the environment 1 events in environ-ment 2 were decoded as moving toward the corresponding position ofthe animal in environment 1. Similarly, for the 24 events in which bothtracks were replayed simultaneously, the direction of the pairs ofenvironment 1/environment 2 replay events was not significantlycorrelated (R ¼ 0.34, P 4 0.05). None of these effects could beattributed to clustering errors or cluster instability (see SupplementaryResults and Supplementary Clusters).

Replay in the final and initial rest sessions

Following exposure to environment 2, rats were placed in the rest boxone last time. Here, we observed statistically significant awake replay(P o 0.05) of both environments, with 51 of 147 (34.7%) beingsignificant candidate environment 1 events and 58 of 128 (45.3%)being significant candidate environment 2 events. Once again, envir-onment 1 and environment 2 appeared to be replayed independently(see Supplementary Results). The decline in the proportion ofenvironment 1 replay events as compared with the previous rest wassignificant (Z¼ 2.07, Po 0.04). There was, however, no correspondingdecline in the pair-wise awake correlation measure applied to cellswith environment 1, but not environment 2, place fields (Rest 4:R2 ¼ 0.1129, 14,631 spike pairs; Rest 2 and 3: R2 ¼ 0.1164, 70,803spike pairs; Z ¼ 0.50, P 4 0.1, not significant). As in the previous restsessions, the probability of replay and the quality of the pair-wisesequential activity was significantly lower during quiescence(5-s immobility criterion; environment 1 replay: 87 of 389 events(22.4%) o 34.7% awake, Z ¼ 2.91, P o 0.01; quiescent: R2 ¼ 0.0604,P o 10�10 compared with awake; environment 2 replay: 105 of 303(34.7%) o 45.3% awake, Z ¼ 2.09, P o 0.05).

Although replay continued into the final rest, we observed minimalreplay activity for either environment during the first rest sessionof the day. The pair-wise regression yielded a significantly lowerR2 value than for all subsequent behavioral sessions (regressionR2 ¼ 0.025; compared with quiescence in rest box, Z¼ 8.30; comparedwith awake periods in rest box, Z ¼ 15.25; compared with runsin environment 2, Z ¼ 19.13; P 4 10�10; Supplementary Table 1

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Figure 3 Replay of environment 1 in the rest box

is more robust during awake than quiescent

periods. (a) The proportion of significant replay

events was similar in environment 1 and during

awake periods in the rest box (R awake) but was

lower during quiescence in the rest box (R quiesc,

*** P o 0.001). (b) Scatter plot of the firing rate

of all neurons with place fields in environment 1during awake and quiescent SWRs in the rest

box. Rates for the large majority of neurons

were higher during quiescence (P o 10�10).

(c) Histogram of the proportions of SWRs during

awake and quiescent periods with different

numbers of active cells. Awake SWRs were more

likely to activate a larger number of cells than

quiescent SWRs (P o 10�10). (d) Pair-wise

reactivation in the rest box. Each plot shows rows

representing the normalized cross-correlegrams

between all pairs of simultaneously recorded

neurons with place fields in environment 1, with

the vertical location of each row being determined by the distance between the two cells place field peaks in environment 1. The V shape representing

activation consistent with replay was more clearly visible in the awake state, and the R2 value representing the degree to which the times between spikes from

two neurons predict the distances between the peaks of their place fields was significantly larger for awake replay events (P o 10�10).

0 0.4 0.8 1.2 1.6 20

0.4

0.8

1.2

1.6

2.0

0

0.1

0.2

0.3

0.4

5 10 15 20Firing rate during awake SWRs (Hz)

Firi

ng r

ate

durin

gqu

iesc

SW

Rs

(Hz)

Number of active cells

Pro

port

ion

of e

vent

s

AwakeQuiescence

0

E1

R awak

e

R quie

sc

10

20

30

40

50

0.5 –0.5 0.50

Rest box (quiescence)

Relative spike timing (s)

Dis

tanc

e be

twee

nfie

ld p

eaks

(cm

)

0

160

80

–0.5 0

Rest box (awake)

Sig

nific

ant e

vent

s (%

)

***a b c

d

4 ADVANCE ONLINE PUBLICATION NATURE NEUROSCIENCE

ART ICLES

and Supplementary Fig. 11 online). Thus, the strength of replayappeared to decay from the end of one day to the next.

Interaction between local spatial input and remote replay

Up until this point, our analysis indicated that environment 1 andenvironment 2 replay were independent. But could spatial informationfrom the local environment contribute to the initiation of remotereplay, as it does for local replay19? If so, cells that receive strongerspatial inputs at a location might be more depolarized and thus morelikely to initiate a replay event. To examine this possibility, we took eachsignificant environment 1 replay event that was seen in environment 2and identified the cells that fired the first and the last spike of the event.We calculated the local spatial firing rate in environment 2 for bothcells, excluding activity in SWRs (see Online Methods).

We found that, on average, the first cell had a higher local firing ratethan the last cell (rank sum, P o 0.002). We repeated this analysis forthe final rest session and found that local rates were higher for the first

cell of awake (Po 0.05), but not quiescent, environment 1 replay events(Fig. 5). A similar trend was present for both Rest 2 and Rest 3 (Supple-mentary Fig. 12 online), and local rates were also higher for the first cellof awake, but not quiescent, replays of environment 2 in the rest box(Supplementary Fig. 13 online). This effect did not result from outlierspikes or other potential confounds (Supplementary Results).

DISCUSSION

We observed coherent spatial replay during SWRs in three differentconditions: awake replay of the currently experienced environment,awake replay of a remote environment and quiescent replay of a remoteenvironment. Awake, remote replay continued long after the initialexperience and represented a more precise recapitulation of pastsequences than replay seen during quiescence. Furthermore, awake,remote replay was very prevalent and the structure of the remote placefields could explain as much as 17% of the variance of SWR spiketiming, indicating that activity consistent with replay is present in asubstantial fraction of SWRs. These results pose difficulties for modelspositing that the ordering of spikes in an awake SWR eventstems from ordered activation of sensory-driven, subthreshold placefields17–19. Our data are instead consistent with a model in which replayactivates representations of previous experiences during both wakingand sleep29.

We found that the initiation of an awake replay sequence was oftenrelated to local spatial input at the rat’s location, but the mnemoniccontent of awake replay could be effectively independent of location.Thus, local inputs could lead to the activation of a set of neurons thatthen activate other associated neurons that are part of a previously storedmemory. We also saw stronger sequential reactivation in CA3 than inCA1, and as SWRs generally originate in CA3 (ref. 10), replay mayinvolve activation of a sequence stored in CA3 associative connections.

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

–0.5 0.500

160

80

Relative spike timing (s)

0

0.2

0.4

E2location

E1

plac

ece

llsE

2 pl

ace

cells

0

0.1

0.2

0

0

20

20

40

40

60

60

80

80

100

100

120

120

140

140

160

160

180

180

200

2000

0.1

0.2

E2location

E1

plac

e c

ells

E2

plac

e c

ells

Decoded distance to center endpoint (cm)

Decoded distance to left endpoint (cm)

Pro

babi

lity

P

roba

bilit

y

Pro

babi

lity

P

roba

bilit

y

f

a

g

e

k

m

0

0.2

0.4

E1replay

E1replay

020406080100120140160180

020406080100120140160180

b

h

c

i

l

d

j

100 ms

100 ms

Dis

tanc

e be

twee

nfie

ld p

eaks

(cm

)

0

0.5

1.5

1.0

2.0

Loca

l spa

tial

rat

e (H

z)

E1 replay event cell order

Awake, rest boxQuiescent, rest box

E2

First FirstLast Last

**

*

n.s.

First Last

Figure 5 Local spatial rate and replay initiation for environment 1 replay

in environment 2 and the rest box. The first cell active in each SWR had a

higher local spatial rate in environment 2 and for awake (o5 s immobile),

but not quiescent (45 s immobile), events in the rest box. Error bars

represents the mean ± 25th percentiles. * P o 0.05 and ** P o 0.002.

n.s., not significant (P 4 0.05).

Figure 4 Robust replay of environment 1 while

the rat was located in environment 2. (a) The

location of the rat in environment 2 during an

environment 1 replay event. (b–d) Spiking during

a SWR representing replay that decoded (c) to a

coherent trajectory in environment 1 (d) (see

Fig. 2c–e for a detailed description of each

element of the plots). For this event, the rat wasstill for 45 s before and after the SWR, so only

the rat’s location is shown in a. (e,f) Activation of

neurons with place fields in environment 2 during

the same SWR event and the decoded locations

for environment 2. The neural activity during the

SWR involved a coherent replay of environment 1,

but not of environment 2. Cells are not numbered

because different subsets of cells are shown in

each panel, but cells with place fields in both

environment 1 and environment 2 are shown in

both raster plots (see Supplementary Fig. 5 for

this event and the associated two-dimensional

spatial rate maps). (g–l) A second example

illustrating replay of environment 1 in the absence

of replay of environment 2 (see Supplementary

Fig. 6 for this event and the associated two-

dimensional spatial rate maps; see Supplementary

Clusters for the cluster plots associated with each

cell with a place field in environment 1). (m) Pair-wise sequential activation plot for environment 2

SWRs including only neurons with place fields in

environment 1, but not in environment 2 (see

Fig. 3d for an explanation of the plot). The

associated R2 value was 0.17, indicating activity

that was consistent with replay.

NATURE NEUROSCIENCE ADVANCE ONLINE PUBLICATION 5

ART ICLES

What might lead to remote rather than local replay? We and othershave found that novel experiences generate a long-lasting increase inneuronal excitability and neuronal coordination for the cells activeduring those experiences13,17,20,30. That increase could contribute tostrong remote replay, as cells representing recent novel experiencewould be easier to activate and would tend to have stronger connec-tions among themselves. In our study, environment 1 was alwaysmore novel than environment 2. This may explain why awake,remote replay of environment 1 was present in as many as 44% ofthe candidate events in environment 2 and was as prevalent as localreplay of environment 2.

The replay events that we observed were present in both waking andmore sleep-like states, but were of higher fidelity in the awake statewhen the rat had recently been in motion. This came as a surprise to us,given that previous results have suggested either similar fidelity ofwaking and sleep SWR reactivation12 or somewhat better reactivationin sleep as compared with waking31. This apparent disparity may reflectthe use of nonsequential coactivity measures in previous studies.

Our findings complement recent reports of memory-related activityin the human32 and rat33 hippocampus. These studies reported that,during waking behavior, the hippocampus can express internallygenerated patterns related to previous experience. Those patternswere present across times scales similar to those seen during experience,however, and were thus distinct from the rapid replay of extendedspatial sequences reported here. We speculate that the higher-rateawake replay that we found could support our ability to retrievememory sequences in much less time than was required for theinitial experience.

Finally, although the causal link between replay and consolidationremains to be established, our results may have important implicationsfor the processes that allow for the long-term storage of new spatial andevent memories. We know that the neocortex and hippocampus canboth show oscillatory, synchronized patterns of activity during sleepand sleep-like states15,34,35. Our findings suggest that this coordinationinvolves a higher rate neural activity, but a less faithful recapitulation ofexperience, than waking events. Awake replay, in contrast, probablyoccurs during a desynchronized neocortical state that is more closelyassociated with sensory processing35. Awake replay may thereforelead to repeated and accurate recapitulations of recent experiencesin neocortical networks and thus more faithful memories forthose experiences. Furthermore, awake replay can reactivate amemory for a past experience in the midst of an ongoing experience,potentially facilitating the formation of associations that linkmultiple distinct events across long spans of time.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank M. Brainard, D. Feldman, S. Lisberger, M. Stryker and the membersof the Frank laboratory for comments on the manuscript. We also thankT. Davidson for a helpful discussion and M. Carr for help with histologyphotographs. This work was supported by the John Merck Scholars Program,the McKnight Scholars Program and US National Institutes of Health grantsMH077970 and MH080283.

AUTHOR CONTRIBUTIONSM.P.K. and L.M.F. designed the experimental procedure. M.P.K. carried out all ofthe data collection and the majority of the analyses. L.M.F. carried out theremaining analyses. M.P.K. and L.M.F. wrote the manuscript.

Published online at http://www.nature.com/natureneuroscience/

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/

1. Scoville, W.B. & Milner, B. Loss of recent memory after bilateral hippocampal lesions.J. Neurol. Neurosurg. Psychiatry 20, 11–21 (1957).

2. O’Keefe, J. & Nadel, L. The Hippocampus as a Cognitive Map (Oxford University Press,London, 1978).

3. Buzsaki, G. The hippocampo-neocortical dialogue. Cereb. Cortex 6, 81–92 (1996).4. Sutherland, G.R. & McNaughton, B. Memory trace reactivation in hippocampal and

neocortical neuronal ensembles. Curr. Opin. Neurobiol. 10, 180–186 (2000).5. Suzuki, W.A. Encoding new episodes and making them stick.Neuron50, 19–21 (2006).6. Wilson, M.A. & McNaughton, B.L. Reactivation of hippocampal ensemble memories

during sleep. Science 265, 676–679 (1994).7. Buzsaki, G., Horvath, Z., Urioste, R., Hetke, J. & Wise, K. High-frequency network

oscillation in the hippocampus. Science 256, 1025–1027 (1992).8. Chrobak, J.J. & Buzsaki, G. High-frequency oscillations in the output networks of the

hippocampal- entorhinal axis of the freely behaving rat. J. Neurosci. 16, 3056–3066(1996).

9. Buzsaki, G. Hippocampal sharp waves: their origin and significance. Brain Res. 398,242–252 (1986).

10. Csicsvari, J., Hirase, H., Mamiya, A. & Buzsaki, G. Ensemble patterns of hippocampalCA3-CA1 neurons during sharp wave–associated population events. Neuron 28,585–594 (2000).

11. Pavlides, C. & Winson, J. Influences of hippocampal place cell firing in the awake stateon the activity of these cells during subsequent sleep episodes. J. Neurosci. 9,2907–2918 (1989).

12. Kudrimoti, H.S., Barnes, C.A. & McNaughton, B.L. Reactivation of hippocampal cellassemblies: effects of behavioral state, experience and EEG dynamics. J. Neurosci. 19,4090–4101 (1999).

13. O’Neill, J., Senior, T.J., Allen, K., Huxter, J.R. & Csicsvari, J. Reactivation of experience-dependent cell assembly patterns in the hippocampus. Nat. Neurosci. 11, 209–215(2008).

14. Lee, A.K. & Wilson, M.A. Memory of sequential experience in the hippocampus duringslow wave sleep. Neuron 36, 1183–1194 (2002).

15. Ji, D. & Wilson, M.A. Coordinated memory replay in the visual cortex and hippocampusduring sleep. Nat. Neurosci. 10, 100–107 (2007).

16. Nadasdy, Z., Hirase, H., Czurko, A., Csicsvari, J. & Buzsaki, G. Replay and timecompression of recurring spike sequences in the hippocampus. J. Neurosci. 19,9497–9507 (1999).

17. Foster, D.J. & Wilson, M.A. Reverse replay of behavioral sequences in hippocampal placecells during the awake state. Nature 440, 680–683 (2006).

18. Diba, K. & Buzsaki, G. Forward and reverse hippocampal place-cell sequences duringripples. Nat. Neurosci. 10, 1241–1242 (2007).

19. Csicsvari, J., O’Neill, J., Allen, K. & Senior, T. Place-selective firing contributes to thereverse-order reactivation of CA1 pyramidal cells during sharp waves in open-fieldexploration. Eur. J. Neurosci. 26, 704–716 (2007).

20. Karlsson, M.P. & Frank, L.M. Network dynamics underlying the formation of sparse, in-formative representations in the hippocampus. J. Neurosci. 28, 14271–14281 (2008).

21. Frank, L.M., Brown, E.N. & Wilson, M.A. Trajectory encoding in the hippocampus andentorhinal cortex. Neuron 27, 169–178 (2000).

22. Kim, S.M. & Frank, L.M. Hippocampal lesions impair rapid learning of a continuousalternation task. PLoS One 4, e5494 (2009).

23. O’Keefe, J. & Recce, M.L. Phase relationship between hippocampal place units and theEEG theta rhythm. Hippocampus 3, 317–330 (1993).

24. Brown, E.N., Frank, L.M., Tang, D., Quirk, M.C. & Wilson, M.A. A statistical paradigm forneural spike train decoding applied to position prediction from ensemble firing patternsof rat hippocampal place cells. J. Neurosci. 18, 7411–7425 (1998).

25. Zhang, K., Ginzburg, I., McNaughton, B.L. & Sejnowski, T.J. Interpreting neuronalpopulation activity by reconstruction: unified framework with application to hippocam-pal place cells. J. Neurophysiol. 79, 1017–1044 (1998).

26. Johnson, A. & Redish, A.D. Neural ensembles in CA3 transiently encode paths forward ofthe animal at a decision point. J. Neurosci. 27, 12176–12189 (2007).

27. Frank, L.M., Stanley, G.B. & Brown, E.N. Hippocampal plasticity across multiple days ofexposure to novel environments. J. Neurosci. 24, 7681–7689 (2004).

28. Euston, D.R., Tatsuno, M. & McNaughton, B.L. Fast-forward playback of recent memorysequences in prefrontal cortex during sleep. Science 318, 1147–1150 (2007).

29. de Hoz, L. & Wood, E.R. Dissociating the past from the present in the activity of placecells. Hippocampus 16, 704–715 (2006).

30. Cheng, S. & Frank, L.M. New experiences enhance coordinated neural activity in thehippocampus. Neuron 57, 303–313 (2008).

31. O’Neill, J., Senior, T. & Csicsvari, J. Place-selective firing of CA1 pyramidal cells duringsharp wave/ripple network patterns in exploratory behavior. Neuron 49, 143–155(2006).

32. Gelbard-Sagiv, H., Mukamel, R., Harel, M., Malach, R. & Fried, I. Internally generatedreactivation of single neurons in human hippocampus during free recall. Science 322,96–101 (2008).

33. Pastalkova, E., Itskov, V., Amarasingham, A. & Buzsaki, G. Internally generatedcell assembly sequences in the rat hippocampus. Science 321, 1322–1327 (2008).

34. Buzsaki, G. & Draguhn, A. Neuronal oscillations in cortical networks. Science 304,1926–1929 (2004).

35. Mehta, M.R. Cortico-hippocampal interaction during up-down states and memoryconsolidation. Nat. Neurosci. 10, 13–15 (2007).

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

6 ADVANCE ONLINE PUBLICATION NATURE NEUROSCIENCE

ART ICLES

ONLINE METHODSA distinct set of analyses of the data used in this study and the associated

methods have been presented previously20.

Data collection and preprocessing. We food-deprived three male Long-Evans

rats (500–600 g) and pretrained them to alternate in a linear track. This

pretraining was performed in a different room from the recording experiments.

After the rats alternated reliably for liquid reward (sweetened condensed milk),

they were implanted with a microdrive array containing 30 independently

movable tetrodes targeting CA3 and CA1 (refs. 20,27) according to University

of California San Francisco Institutional Animal Care and Use Committee and

US National Institutes of Health guidelines. On the days following surgery, the

tetrodes were advanced to the cell layers. All neural signals were recorded

relative to a reference tetrode in the corpus callosum. Following data collection,

electrode locations were identified histologically (see Supplementary Fig. 14

online for example histological sections and a diagram of the targeted regions

of CA1 and CA3).

Rats were introduced to W track environment 2 (environment 2) either 3

(n ¼ 1) or 6 (n ¼ 2) d before being introduced to W track environment 1

(environment 1), with two 15-min run sessions per day. We examined data

from when the rats were first introduced to environment 1 onward, when the

rats ran two sessions in environment 1 followed by one session in environment

2. The rats were rewarded for performing a continuous alternation task20,21.

Rapid learning in this task requires an intact hippocampus22. The rats

performed the inbound portion of the task (left or right to center) at nearly

100% correct on all days and performed at above chance levels on the

outbound portion of the task (center to left or right) beginning with the

second day of exposure to environment 1.

Run sessions were flanked by 20-min rest periods in a high-walled black

box (floor, 25 � 34 cm; walls, 50 cm tall). The W tracks were 76 � 76 cm

with 7-cm-wide track sections. Tetrode positions were adjusted after daily

recording sessions for all tetrodes that had poor unit recordings. On rare

occasions, some tetrodes were moved before recording sessions but never

within 4 h of recording.

Data were collected using the NSpike data acquisition system (L.M.F. and

J. MacArthur, Harvard Instrumentation Design Laboratory). An infrared diode

array with a large and a small cluster of diodes was attached to the preamps

during recording. Following recording, the rat’s position on the track was

reconstructed using a semi-automated analysis of digital video of the experi-

ment. Individual units (putative single neurons) were identified by clustering

spikes using peak amplitude and spike width as variables (MatClust, M.P.K.).

Care was taken to only cluster well-isolated neurons with spike waveform

amplitudes that were clearly stable over the course of the entire session. We

were frequently able to use a single set of cluster bounds defined in amplitude

and width space to isolate units across an entire 2–3-h recording session. In the

minority of cases in which there was a slight shift in amplitudes across time,

units were clustered only when that shift was coherent across multiple clusters

and when plots of amplitude versus time showed a smooth shift. Thus, no units

were clustered in which part of the cluster was in the noise or was cut off at the

recording threshold.

Analysis of neural data. Analyses were carried out using custom software

written in Matlab (Mathworks). To measure place field locations, we calculated

the linearized activity of each cell. The rat’s linear position was measured as the

distance in centimeters along the track from the reward site on the center arm.

We then produced an occupancy-normalized firing-rate map using spike

counts and occupancies calculated in 2-cm bins and smoothed with a 4-cm

s.d. Gaussian curve. Only times outside of SWRs (see below) were included.

The place field peak rate was defined as the maximum rate across all spatial

bins. A peak rate of 3 Hz or greater was required for a cell to be considered a

place cell. The results were the same with the 5-Hz threshold that was used in

previous reports17,18. Putative interneurons were identified on the basis of

spike width and average firing rate36,37 and were excluded from all analyses.

Two-dimensional occupancy-normalized spatial rate maps (Fig. 1 and

Supplementary Figs. 6, 7 and 10) were constructed with 2-cm square bins

of spike count and occupancy, both smoothed with a two-dimensional

Gaussian (8-cm s.d.). Once again, all SWRs were excluded.

SWRs were identified on the basis of peaks in the LFP recorded from one

channel from each tetrode in the CA3 and CA1 cell layers. The raw LFP data

were bandpass-filtered between 150–250 Hz, and the SWR envelope was

determined using a Hilbert transform. The envelope was smoothed with a

Gaussian (4-ms s.d.). We initially identified SWR events as sets of times when

the smoothed envelope stayed above 3 s.d. of the mean for at least 15 ms on at

least one tetrode. We defined the entire SWR as including times immediately

before and after that threshold crossing event during which the envelope

exceeded the mean. Overlapping SWRs were combined across tetrodes, so

many events extended beyond the SWR seen on a single tetrode.

We similarly extract theta and delta activity from the bandpass–filtered LFP

(theta, 6–12 Hz; delta, 0.5–4 Hz). We used a Hilbert transform to determine the

envelope of both theta and delta, smoothed both envelopes with a Gaussian

(1-s s.d.) and computed the theta/delta ratio by dividing the magnitude of the

smoothed theta envelope by the magnitude of the smoothed delta envelope. For

each day, we selected the CA3 tetrodes where the variance of the theta/delta

ratio was highest and normalized the values to a mean of 1 to permit

combining data across days.

Candidate replay events were defined as SWRs during which at least five

place cells from the replayed environment fired at least one spike each. We

determined the sequential representation of position seen during a candidate

replay using a simple Bayesian decoder24,25. Each event was divided into 15-ms

bins, and for each bin with at least one spike in it, we calculated the spatial

probability distribution

PðX jNC1 Þ ¼ PðNC

1 jXÞPðXÞPðNC

1 Þ

where X is the set of all locations in the environment (using 2-cm bins) and NC1

is a vector of spike counts for all cells (C) that had place fields in the

environment. Bins without spikes cannot be decoded using this simple

algorithm, so these bins were omitted from the analysis. PðNC1 jXÞ was

calculated using the approximation that different cells are independent,

PðNC1 jXÞ ¼

QC

i¼1

PðNi jXÞ. PðNi jXÞ is the probability, at each location

in the track, that cell i fired Ni spikes. We estimated the probability on the basis

of the occupancy-normalized rate maps for that cell and the assumption that

spike counts are Poisson distributed. P(X) was a uniform distribution across all

spatial bins and was thus an uninformative prior. PðNC1 Þ was not estimated.

Instead, PðX jNC1 Þ was normalized across X to sum to 1. Note that decoding

was done between food well locations, but as the rat could move slightly

beyond these locations, the two-dimensional place field plots extend further

than the linear distances used for decoding.

To determine whether a given decoded sequence was unlikely to occur by

chance, we drew 10,000 random samples from the PðX jNC1 Þ distribution for

each decoded bin and assigned the sampled locations to that bin. We then

performed a linear regression on the bin number versus location points. The

resulting R2 value was then compared to 10,000 regressions in which the order

of the bins was shuffled. This shuffling preserved the spiking structure in each

bin, but ordered the bins randomly. This is equivalent to shuffling the order of

the probability distributions over position produced from the decoding analysis

(Supplementary Fig. 15 online). The P value for the event was the proportion

of the shuffled R2 values that was greater than the R2 value of the actual event,

and an event with P o 0.05 was considered to be unlikely to occur by chance.

We obtained similar results when we shuffled the identity of each cell (for

example, assigned spike trains to randomly chosen neurons18; data not shown).

We also developed a pair-wise measure of sequential activity that was

consistent with replay. For every pair of cells, we measured the absolute value

of the time from each reference spike of one cell to all spikes from the other cell.

Only spikes occurring during SWRs and only times up to 500 ms were

included. We also measured the linear distance between the place field peaks

of the two neurons as the shortest path on the environment from the peak

location for one cell to the peak location for the other cell. We then used a

linear regression to calculate the R2 value which measured the degree to which

the distance between two cells’ place fields predicted the absolute value of the

time between SWR spikes from the cells. Each pair of cells was included

only once.

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

doi:10.1038/nn.2344 NATURE NEUROSCIENCE

We illustrated the quality of pair-wise reactivation using a method similar to

that developed in a previous manuscript28. We computed the cross-correlation

histogram between SWR spikes from all pairs of neurons with place fields in the

environment (5-ms bins, �500 to 500 ms extent). We constructed a two-

dimensional histogram plot in which the time between spikes was on the

x axis, and the counts from each cross correlation histogram were placed in a

row at the distance between the centers of the two place fields. Thus, the

correlogram for a pair of cells with identical place field peak locations would be

a row of values at the bottom of the plot at a y value of 0, whereas the

correlegram from two cells whose fields were 50 cm apart would be a row of

values at a y value of 50. Each row was normalized so that the peak and trough

ranged from 0 to 1.

The similarity of spatial coding for single cells across the two environments

was computed using place field overlap, defined as twice the sum of the

overlapping areas of the linear rate curves divided by the sum of the areas of

each curve38. This measure is bounded between 0 and 1, where 0 signifies no

overlap and 1 signifies perfect overlap. This is equivalent to rotating one of the

tracks so that it lies entirely on top of the other and computing the similarity of

linear place fields on the basis of that rotation.

We determined whether replay of environments 1 and 2 was correlated by

examining all replay events that were simultaneously environment 1 and

environment 2 candidate events. Thus, each event had five or more environ-

ment 1 place fields and five or more environment 2 place fields active. As some

cells were active in both environments, the total number of active cells was in

some cases less than 10. We measured the R2 value for the decoding of both

environment 1 and environment 2 and used the shuffling described above to

produce a P value for both the environment 1 and the environment 2 cells. We

computed the �log10 of the P values and calculated the correlation coefficient

for the environment 1 and environment 2 transformed P values. We also

examined candidate events for both environment 1 and environment 2 and

asked whether the transformed P value of the environment 1 replay was related

to the number of active environment 2 cells.

We examined the effect of local spatial firing on remote replay for replay

occurring both while the rat was in environment 2 and while it was in the rest

box. For each significant replay event, we identified the cell that fired the first

spike and the cell that fired the last spike of the sequence. For environment 2,

we used the linear place fields generated for the decoding to determine the rate

of each cell at the rat’s location. As indicated above, these place fields did not

include SWR activity. For the rest box, we similarly excluded SWRs and

computed a two-dimensional firing rate in 2-cm bins following smoothing

with a two-dimensional Gaussian (4 cm s.d.). We then calculated the mean

rates for the first and last cells across all significant events.

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

36. Frank, L.M., Brown, E.N. & Wilson, M.A. A comparison of the firing properties of putativeexcitatory and inhibitory neurons from CA1 and the entorhinal cortex. J. Neurophysiol.86, 2029–2040 (2001).

37. Fox, S.E. & Ranck, J.B.J. Electrophysiological characteristics of hippocampal complex–spike cells and theta cells. Exp. Brain Res. 41, 399–410 (1981).

38. Battaglia, F.P., Sutherland, G.R. & McNaughton, B.L. Local sensory cues and place celldirectionality: additional evidence of prospective coding in the hippocampus.J. Neurosci. 24, 4541–4550 (2004).

NATURE NEUROSCIENCE doi:10.1038/nn.2344