Human brain activity and functional connectivity as memories age from one hour to one month

Abstract

Theories of memory consolidation suggest the role of brain regions and connectivity between brain regions change as memories age. Human lesion studies indicate memories become hippocampus-independent over years, whereas animal studies suggest this process occurs across relatively short intervals, days to weeks. Human neuroimaging studies suggest that changes in hippocampal and cortical activity and connectivity can be detected over these short intervals, but many of these studies examined only two time periods. We examined memory and FMRI activity for photos of indoor and outdoor scenes across four time periods to examine these neural changes more carefully. Participants (N=21) studied scenes 1 hour, 1 day, 1 week, or 1 month before scanning. During scanning, participants viewed scenes, made old/new recognition memory judgments, and gave confidence ratings. Memory accuracy, confidence ratings, and response times changed with memory age. Brain activity in a widespread cortical network either increased or decreased with memory age, whereas hippocampal activity was not related to memory age. These findings were almost identical when effects of behavioral changes across time periods were minimized. Functional connectivity of the ventromedial prefrontal cortex with the posterior parietal cortex increased with memory age. By contrast, functional connectivity of the hippocampal with the parahippocampal cortex and fusiform gyrus decreased with memory age. In sum, we detected changes in cortical activity and changes in hippocampal and cortical connectivity with memory age across short intervals. These findings provide support for the predictions of systems consolidation and suggest that these changes begin soon after memories are formed.

Introduction

The stabilization of long-term memory is a gradual process dependent on the phenomenon of memory consolidation. After learning, declarative memories are thought to become hippocampus-independent via systems consolidation (Marr, 1971; Squire & Alvarez, 1995). Specifically, structures in the medial temporal lobe (MTL) first encode information and then, via bidirectional connections with neocortex the memory is reorganized such that it can eventually be retrieved without the MTL.

Studies of patients with permanent lesions limited to the hippocampus suggest that systems consolidation occurs over several years, while other MTL structures (the parahippocampal gyrus) may have a longer role in this process (Bayley et al., 2006; Manns et al., 2003). Non-invasive measures of MTL activity (e.g., functional magnetic resonance imaging, fMRI) and measures of the functional connections between brain regions (functional connectivity) also show evidence of changes in brain activity or connectivity across several years. For example, brain activity in the MTL during semantic (fact) memory retrieval decreased as a function of memory age across several years (Douville et al., 2005; Haist, Gore, et al., 2001; Smith & Squire, 2009). Changes in connectivity with the MTL as a function of memory age have not yet been examined for semantic memory, but studies of autobiographical (episodic) memory have documented decreasing hippocampal-cortical connectivity with memory age across several years (Gilmore et al., 2021; Sheldon & Levine, 2013; Söderlund et al., 2012).

In animals with hippocampal lesions, the timescale of systems consolidation is shorter, occurring over days to weeks (Clark et al., 2002; Kim et al., 1995; Kim & Fanselow, 1992; Winocur, 1990; Zola-Morgan & Squire, 1990). Concordant findings of decreases in hippocampal activity in animals also reveal evidence for long-term memory consolidation over this shorter timescale using measures of glucose uptake (Bontempi et al., 1999) or immediate early gene activity (Frankland et al., 2004; Maviel et al., 2004; Wheeler et al., 2013). fMRI studies in humans have investigated consolidation across these short time intervals (from minutes to months) in search of early evidence of memory consolidation. However, unlike the long-term consolidation literature, findings associated with changes in the MTL activity across short time intervals are mixed.

Contrary to systems consolidation theory, multiple-trace theory (MTT), predicts that the hippocampus is always needed for retrieval of episodic memories (Moscovitch et al., 2005; Nadel et al., 2000). Extending the predictions of these theories, which were based largely on findings from patients with hippocampal or MTL lesions, to neuroimaging measures of brain activity and connectivity is complicated by the fact that neuroimaging measures do not speak to whether regions are necessary for the purported functions. Nevertheless, there are predictions from these theories that can be examined using neuroimaging measures and that can provide concordant evidence in studies of long-term memory consolidation over short time intervals. Specifically, hippocampal activity that decreases over time (right hippocampus only, Bosshardt, Schmidt, et al., 2005; Dandolo & Schwabe, 2018; item memory, Du et al., 2019; Furman et al., 2012; Harand et al., 2012; Milton et al., 2011; Ritchey et al., 2015; Sekeres, Winocur, Moscovitch, et al., 2018; delay between visual cue and auditory stimulus, Smith et al., 2010; Sterpenich et al., 2009; Takashima et al., 2009; Takashima et al., 2006; Yamashita et al., 2009), is commonly taken as support of systems consolidation theory, whereas hippocampal activity that increases over time (supposedly reflecting the addition of multiple memory traces within the hippocampus) (Bosshardt, Degonda, et al., 2005; left hippocampus only, Bosshardt, Schmidt, et al., 2005; Gais et al., 2007; visual cue only, Smith et al., 2010), is taken in support of MTT. Note that an absence of significant changes in hippocampal activity across time periods (Davis et al., 2009; associative memory, Du et al., 2019; Janzen et al., 2008; Stark & Squire, 2000; Suchan et al., 2008; Takashima et al., 2017; Tompary & Davachi, 2017; Vilberg & Davachi, 2013), is sometimes taken to support MTT because this theory posits that the hippocampus is always necessary for retrieval. This interpretation is problematic because neuroimaging cannot test whether a region is necessary (as opposed to related or interested, but unnecessary) and because null findings can result from attributes unrelated to consolidation, such as insufficient power or poor signal in the MTL.

There is more consistent support for systems consolidation across short time periods for measures of functional connectivity between the hippocampus and neocortex during memory retrieval (Brodt et al., 2016; Takashima et al., 2009) or reflected as alterations in connectivity during post-encoding rest periods (van Kesteren et al., 2010). Animal studies have also found evidence of changing connectivity between the hippocampus and neocortex with memory age (e.g., Bontempi et al., 1999; Wheeler et al., 2013; Wirt & Hyman, 2019).

It is unclear why the literature on hippocampal activity related to memory consolidation over short time intervals is so mixed. Recent theoretical perspectives of memory consolidation [neural-psychological representation correspondence (NPRC) and contextual binding theory (CBT)] highlight several ways in which changes in brain activity or connectivity with memory age activity may reflect other factors besides memory age (Gilboa & Moscovitch, 2021; Yonelinas et al., 2019). For example, brain activity may simply be tracking the substantial behavioral changes that are observed across short time intervals (i.e., decreases in accuracy and confidence and increases in response time). A considerable number of studies examining consolidation over short time intervals do not account for these behavioral changes when interpreting brain activity/connectivity changes. Some studies have addressed the influence of these behavioral changes by limiting analysis to the trials remembered with high confidence (e.g., high-confidence hits, recollections from Remember-Know procedure) (Dandolo & Schwabe, 2018; Harand et al., 2012; Milton et al., 2011; Ritchey et al., 2015; Sterpenich et al., 2009; Takashima et al., 2009; Takashima et al., 2006; Tompary & Davachi, 2017; Yamashita et al., 2009). While this approach may reduce the impact of behavioral changes, it disproportionally omits trials in remote time periods (which have fewer high-confidence hits due to forgetting) and the behavioral differences often persist even after limiting analysis to these trials. An alternative approach to reduce the impact of these behavioral changes is to use all trials and employ statistical analysis that covaries out trial-by-trial fluctuations in behavior (amplitude modulation analysis).

Another factor that may contribute to the mixed findings is that brain activity likely reflects the competing influences of memory retrieval and memory encoding (or re-encoding), particularly for the MTL (Buckner et al., 2001; Nadel et al., 2000; Smith & Squire, 2009). Specifically, during retrieval of recent and remote memories, participants encode the content of the memory test (and re-encode the studied items). Due to forgetting, weakly remembered items in more remote time periods should experience more re-encoding than the more strongly remembered items in recent time periods. Therefore, changes in brain activation across recent and remote conditions likely reflects a mixture of encoding and retrieval processes and this phenomenon may complicate detection of changes that reflect the age of the memory.

Lastly, the majority of memory consolidation studies over short time intervals examined only two time periods. A significant change in activity across only two time periods in either a decreasing or increasing direction (or a null result) can be taken as support for systems consolidation or MTT, respectively. One way to ameliorate this issue is to examine more than two time periods and identify patterns of activity consistent with memory consolidation. There are only a few studies which have examined three time periods (Gais et al., 2007; Stark & Squire, 2000; Suchan et al., 2008; Yamashita et al., 2009) and two studies that investigated four time periods (Du et al., 2019; Takashima et al., 2006).

Therefore, we examined brain activity and brain connectivity associated with retrieval of scenes learned one hour, one day, one week, or one month earlier (four time periods). We preserved all trials for analysis by using amplitude modulation analyses to minimize the confounding effects of behavioral changes on neuroimaging measures across time periods. In addition, to identify regions important for memory encoding, we obtained measures of the encoding that occurred during the retrieval test. Finally, to contextualize our findings, we carried out a systematic review of similar memory consolidation studies that examine short time intervals to determine which methodological factors appear to influence the patterns of brain activity in the MTL and neocortex.

Methods

Participants

Twenty-one participants (10 female; mean age = 29 ± 1.3 yr; range = 21–42 yr; mean education = 17 yr ± 0.5) were recruited from the San Diego community.

Materials and procedure

These experimental procedures were reported previously (Urgolites et al., 2015) in the context of a report on brain activity for true versus false memory. The present analysis and findings for brain activity and connectivity as a function of the age of memory are orthogonal to this earlier report.

Participants completed a recognition memory task in the fMRI scanner in which they were asked to remember a set of previously presented indoor and outdoor scenes that had been studied either one month, one week, one day, and one hour before scanning. Scenes were chosen for the experiment that minimized the possibility that participants could confuse one scene for another. Participants were encouraged to deeply encode 320 unique scenes (80 scenes per study session) by answering one of the following questions: “Is this an everyday scene?,” “Does the scene remind you of a place you have been?,” and “Can you picture yourself in the scene?”. Participants viewed each scene for 5.5 sec and were instructed to press “yes” or “no” on a keyboard to respond. These encoding questions were randomized across scenes and across participants.

During the scanning session, participants viewed 240 scenes (60 from each study phase), 240 novel scenes, and 606 baseline trials (Figure 1). For scene trials, scenes were presented for 1 sec and participants subsequently had 3 sec to make a recognition memory judgment with confidence ratings (1 = definitely new, 2 = probably new, 3 = maybe new, 4 = maybe old, 5 = probably old, and 6 = definitely old). Recognition judgments were made by selecting a number (1–6) on the screen with an MRI-compatible mouse (Current Designs, Philadelphia, PA). For baseline trials, a single digit (1–8) was presented and participants selected whether it was odd or even with the mouse (2 sec) (Stark & Squire, 2001). Zero to seven baseline trials were presented after each scene trial (mean = 2.5 trials). Before each scene trial, the starting location of the mouse cursor on the screen was randomized with the intention of decorrelating rightward and leftward movements with the right and left sides of the 1–6 scale. The scanning session consisted of six 8.7 min runs, with each run containing 80 scenes (10 target scenes from each study session intermixed with 40 novel foil scenes). After scanning, a surprise post-scan recognition memory test was administered outside of the scanner. Participants were presented with the 240 foil scenes viewed in the scanner (now considered to be targets) intermixed with 240 novel foil scenes. The test was administered across four runs and participants had unlimited time to respond. The same recognition judgement scale was used for participants to respond if the scene was old (previously seen in the scanner), or new (first time encountering the scene). Each scene had an equal chance of being presented during one of the study phases, during the test phase as a target or foil, or during the post-scan recognition memory test as a foil.

Figure 1.

Note. Study design. Participants studied a different set of 80 color photographs of indoor and outdoor scenes one hour, one day, one week, or one month before scanning. In the scanner, participants made old/new recognition memory judgments with confidence ratings (1 = definitely new to 6 = definitely old) in response to scenes (240 studied scenes intermixed with 240 novel scenes) and made odd/even judgments in response to digits (baseline trials).

fMRI imaging

Scanning was conducted on a 3T General Electric scanner at the Center for Functional MRI (University of California San Diego). Functional images were acquired using a gradient- echo, echo-planar, T2*-weighted pulse sequence (2000 msec TR; 64 × 64 matrix size; 25 cm field of view; 3.9 × 3.9 mm in-plane resolution. Thirty-six oblique coronal slices (slice thickness = 4.8 mm) were acquired perpendicular to the long axis of the hippocampus and covering the whole brain. Following the six functional runs, high-resolution structural images were acquired using a T1-weighted IR-SPGR pulse sequence (25.6 cm field of view; 172 slices; 1.0 mm slice thickness; 256 × 256 matrix size).

Data Analysis

Behavioral data analysis

Measures of accuracy (percent correct), confidence, and response time were calculated for each time period by taking the mean across all targets in the time period. Accuracy was calculated using the following formula: [hit rate/(hit rate + false alarm rate)] X 100. Discriminability [d prime (d’)] was also calculated using the following formula: Z(hit rate) – Z(false alarm rate) using Excel. Means and SEM are reported.

fMRI data analysis

Preprocessing

Preprocessing and statistical modeling of fMRI data were carried out using the Analysis of Functional NeuroImages (AFNI) suite of programs (Cox, 1996) and Advanced Normalization Tools (ANTs) (Avants et al., 2011). During preprocessing, functional data were bias field corrected using field maps collected prior to the functional scans. Using afni_proc.py, large spikes in signal were removed from the functional timeseries. Each volume was then slice-time corrected and co-registered with the anatomical scan using the normalized mutual information cost function. Co-registered functional data for each run were transformed to a common template (Kirby atlas; (https://figshare.com/articles/ANTs_ANTsR_Brain_Templates/915436) using the ANTs programs antsRegistrationSyn.sh and antsApplyTransforms.sh. The Kirby template was selected because the average age matched our sample and because it had precise delineation of the hippocampus compared to other standard spaces. The average spatial smoothing after preprocessing was 6.4 mm (measured using 3dFWHMx), therefore no further spatial smoothing was applied. After spatial transformation, each run was masked with a dilated Kirby atlas to eliminate signal outside the brain, scaled so that mean activation for each run was 100 for each voxel, and resampled to 3 mm isotropic voxels. The first five volumes were excluded from analyses to allow for signal equilibration.

Multiple regression analyses (using AFNI’s 3dDeconvolve) were carried out to obtain amplitude measures for vectors of interest. The multiple regression analyses included 12 motion vectors (6 coding for translation and rotation and 6 coding for the derivatives of these vectors), and the vectors of interest (described below). The hemodynamic response function was estimated for each scene trial independently using eight TENT functions spanning 0–14 seconds after the onset of the trial. Peak activation from 4–8 seconds post-stimulus onset was isolated for analysis (the third and fourth TENT functions). In order to remove the influence of temporal autocorrelation on the multiple regression analysis, the deconvolution matrix was fit to the functional data using generalized least square regression within 3dREMLfit. The resulting β coefficients for each vector of interest from the multiple regression model reflect the mean peak activation.

Whole-brain, voxel-wise analysis of brain activity

Primary Analysis: Memory Age.

The main goal was to identify brain regions associated with memory consolidation. We asked which brain regions exhibited retrieval-related activity that changed as a function of memory age across the four time periods (one hour, one day, one week, and one month). We followed the method used by (Takashima et al., 2006) to detect regions where activity followed a power function across the time periods, a pattern that resembles the power law of forgetting (Wickelgren, 1974; Wixted & Carpenter, 2007). First, four vectors of interest were created that coded for the target trials (hits and misses) for each time period. A fifth vector coded for foil trials. Second, multiple regression analysis was used to obtain beta coefficients for each time period. Third, we carried out a group-level analysis to examine differences in the beta coefficients that changed across time periods according to a power function (y= x^−0.33) using a linear mixed effect model (LME) (Chen et al., 2013). Finally, voxels were clustered using a voxel-wise probability value of p < 0.001 and a cluster-wise probability of p < 0.05 (Woo et al., 2014). Patterns of activity that increased or decreased in relatively monotonic patterns across the time periods were of primary interest, because the detection of a pattern of activity across several time periods is likely more robust than detection of differential activity across only two time periods.

Secondary Analyses.

Next, we sought to clarify if the clusters identified by the primary analysis (retrieval-related activity associated with memory age, described above) might instead reflect other factors that changed as a function of memory age. We carried out two secondary analyses to contextualize our primary findings.

Amplitude-modulated analysis:

The amplitude-modulated analysis sought to determine if the brain regions identified by the primary analysis might have been affected by the behavioral changes that were observed across the time periods (Figure 2). Brain activity that followed the behavioral changes could mimic or interfere with the memory consolidation effects from the primary analysis. Specifically, it was possible that the relatively monotonic increases or decreases in brain activity observed in the primary analysis might have instead reflected concomitant changes in memory confidence or response times across the time periods. Alternatively, brain activity related to the behavioral changes could have masked brain activity associated with memory age. To minimize the effects of these behavioral changes on brain activity, we carried out the same multiple regression analysis described for the primary analysis, but with two additional vectors for each vector of interest. For each of the five vectors in the primary analysis (non-amplitude modulated vectors), two new vectors were created that coded for trial-by-trial variability in memory confidence or response times, respectively (10 additional vectors; 3dMarry).

Figure 2.

Note. In-scanner behavioral measures of memory. Average in-scanner recognition memory performance for all trials as a function of memory age. A. Measure of discrimination (d’ [D Prime]) between old and new scenes B. Confidence ratings from 1 (definitely new) to 6 (definitely old) for targets. C. Response time in milliseconds for targets. Error bars show SEM.

Because the amplitude-modulated vectors and the non-amplitude modulated vectors were included in the multiple regression analysis together, the resultant beta coefficients associated with the non-amplitude modulated vectors now reflected retrieval-related activity when the effects of the behavioral changes were minimized. The same LME analysis was carried out as described for the primary analysis, except that the beta coefficients for the vectors of interest were taken from the amplitude-modulated multiple regression analysis.

In order to identify whether these trial-by-trial measures of confidence and response time were indeed coding for these behavioral effects, we also carried out analyses to identify these networks. Specifically, we conducted one-sample t-tests on the vectors associated with confidence for each time period and the vectors associated with response time for each time period to identify brain activity that corresponded to these trial-by-trial changes. We used the technique Equitable Thresholding and Clustering (ETAC) to ensure a false positive rate of < 1% (Cox, 2019). This method uses random permutation testing and allows for greater accumulation of statistical evidence and the simultaneous use of multiple per-voxel p-value thresholds while controlling the false positive rate. We also tested whether the association between brain activity and these trial-by-trial variations changed across time periods using LME, using the same approach described in the primary analysis.

Due to difficulties encountered during data collection, response times were unavailable for one participant and confidence ratings were unavailable for two participants. For these three participants, the multiple regression analysis included only one amplitude-modulated regressor for each vector of interest.

Note that prior studies have attempted to minimize the effects of behavioral changes across time periods by selecting trials that were correctly remembered (hits) with high confidence (Takashima et al., 2009; Takashima et al., 2006; Yamashita et al., 2009) and carrying out analyses of only these trials. Because memory accuracy and confidence typically decrease as time passes after learning, this method results in exclusion of many trials from analysis and more trials are excluded from remote time periods than recent time periods. In addition, even in these select trials, differences in confidence and response times can still be detected. Accordingly, we carried out the amplitude-modulated analysis to minimize the effects of behavioral changes because it allowed for examination of all trials and equal weighting of trials for all time periods.

Re-encoding analysis:

Another factor that may have affected retrieval-related activity associated with memory age is the level of re-encoding associated with the targets in each of the time periods. This pattern of activity could mimic or interfere with the memory consolidation effects from the primary analysis. Because forgetting occurred across the time periods (Figure 2), weakly remembered targets in the more remote time periods may have experienced more re-encoding than more strongly remembered targets in the more recent time periods. For example, brain regions involved in encoding (and not retrieval) might exhibit increased activity as a function of memory age. To identify brain regions involved in encoding, we examined activity associated with incidental encoding of the foils during the in-scanner memory retrieval test. Specifically, we created behavioral vectors for foils presented during scanning that were later remembered (successful encoding) or forgotten (unsuccessful encoding) according to subsequent memory for the foils from the post-test. There were 184 ± 6.8 remembered foil trials and 56 ± 6.8 forgotten foil trials per participant. A multiple regression analysis was carried out that included these vectors of interest (remembered foil trials or forgotten foil trials) as well as a vector that coded for all target trials. A t-test was used to detect brain regions that distinguished remembered from forgotten foils, ETAC threshold of a false positive rate of < 1%.

Region of interest analysis: Hippocampus

Given the time-limited role of the hippocampus in systems consolidation, we carried out two follow-up analyses of activity in the hippocampus using priori regions of interest (ROI). First, an anatomical ROI was created from manual tracings (bilaterally) drawn in native space on the T1-weighted images for 14 of the 21 participants. This ROI was warped to the study-specific template from these participants from Urgolites et al. (2015). The hippocampal ROI from the study-specific template was then warped to Kirby space for all participants. The bilateral hippocampal ROI was additionally separated into the left and right sides and anterior and posterior components.

Second, a functional hippocampal ROI was created by contrasting activity associated with remembered targets (hits) versus forgotten targets (misses). We created behavioral vectors for all hit trials (regardless of time period), all miss trials (regardless of time period), and all foil trials. A multiple regression analysis was carried out that included these vectors of interest in order to obtain beta coefficients for these trial types. A t-test was used to detect clusters where activity distinguished remembered targets (hits) from forgotten targets (misses) at a threshold with a false positive rate of <1% using ETAC. One large cluster was detected in the MTL bilaterally that overlapped with the hippocampus. A functional ROI was created by selecting voxels in the hippocampus from the significant clusters and the resulting ROI had 87 voxels (2349 mm³). We were aware that carrying out an analysis of brain activity for targets as a function of time period in a functional ROI obtained from the same data (target hits versus target misses) is a circular analysis (Kriegeskorte et al., 2009) and this method is associated with an increased likelihood of type 1 error. In this case, the analysis was carried out to ascertain if the lack of significant findings from the whole-brain analysis and the anatomical ROI analysis would persist even for an analysis that was well-positioned to find a significant result.

For the anatomical ROI and functional hippocampal ROI, the mean beta coefficients from the primary analysis (i.e., for each time period) were averaged across all voxels within the ROIs. A linear mixed effect model was used to determine if brain activity followed a power function and a repeated-measures ANOVA (Memory age: hour, day, week, month) was used to determine if activity changed as a function of memory age.

Whole-brain, voxel-wise analysis of functional connectivity

The main goal was to identify brain regions where functional connectivity changed as a function of memory age. Connectivity was examined using a generalized psychophysiological interaction approach (gPPI) (Friston et al., 1997; McLaren et al., 2012) where activity in a seed region is used to predict activity elsewhere in the brain. Although gPPI does not explicitly test for directionality, it conveys the influence of the seed on other brain regions during a task condition (Gerchen et al., 2014), thus better characterizing functional relationships than simply correlating activity between regions. Two separate gPPI analyses were carried out to assess the influence of a priori seed regions (ROIs) on voxels in the rest of the brain. The ROIs were the bilateral hippocampus and the bilateral ventromedial prefrontal cortex (vmPFC). The hippocampal seed region was the anatomical ROI described above. The vmPFC seed region was derived from Takashima et al. (2006) where activity in their vmPFC cluster increased with memory age from one day to three months. Specifically, based on the MNI coordinate of their cluster (−2, 32, 10), a spherical ROI was created with a radius of 8 mm. For each gPPI analysis, 3dSynthesize was used to extract the baseline activity from the primary regression analysis of memory age (described above). Baseline activity was then subtracted from the preprocessed, concatenated functional runs using 3dcalc. Neural interaction vectors were created for trials associated with each condition of interest. These vectors reflected neural activity in the seed region convolved with trial onset times for the conditions of interest (foils, one hour, one day, one week, and one month time periods). The neural-interaction vectors were added to the multiple regression for the primary analysis of memory age and beta coefficients were obtained for each neural interaction vector. These beta coefficients represent the amplitude of the influence of the seed region on voxels in the brain for each condition of interest. Because the neural interaction vectors of interest and the original vectors of interest were included in the same regression model, the beta coefficients obtained from the neural interaction vectors represent effects that are independent of the brain activity effects. The resulting beta coefficients associated with each time period’s neural-interaction vector were analyzed using the same LME analysis as described for the brain activity analysis.

Results

Behavioral findings

During the recognition memory test in the scanner, accuracy, old-new discriminability (d’), and confidence ratings decreased while response time increased across time periods (Figure 2). Participants obtained mean accuracy (percent correct) scores of 78.17 ± 1.55% for the hour condition, 76.39 ± 2.05% for the day condition, 72.74 ± 2.82% for the week condition, and 71.04 ± 1.72% for the month condition (F(_3,54)=6.09, η² = 0.253, p < 0.001). Participants obtained old-new discriminability scores of 2.28 ± 0.09 for the hour condition, 2.01 ± 0.16 for the day condition, 1.48 ± 0.17 for the week condition, and 1.15 ± 0.15 for the month condition (F(_3,54)=16.51, η² = 0.478, p < 0.001). The average response times (ms) were 1280.20 ± 56.65 for the hour condition, 1392.34 ± 62.32 for the day condition, 1521.63 ± 52.75 for the week condition, 1527.00 ± 62.91 for the month condition (F(_3,57) = 17.71, η² = 0.482, p < 0.001). Finally, the average confidence scores were 5.81 ± 0.03 for the hour condition, 5.58 ± 0.08 for the day condition, 5.29 ± 0.10 for the week condition, and 5.12 ± 0.09 for the month condition (F(_3,54) = 21.52, η² = 0.545, p < 0.001). On the surprise post-scan recognition memory test, participants exhibited 80.34 ± 1.50% correct and obtained a d’ of 1.91 ± 0.12.

Brain regions where activity changed as a function of memory age

The primary analysis detected a network of 18 brain regions where activity changed with memory age including the prefrontal cortex (PFC), sensorimotor cortex, posterior parietal cortex (PPC), and right middle occipital gyrus (Figure 3, Table 1). All regions demonstrated relatively monotonic increases or decreases in activity as a function of memory age. Notably, no regions in the MTL were detected even when the probability level was lowered substantially.

Figure 3.

Note. Brain regions in which activity followed memory age. A. Coronal sections displaying brain regions in which brain activity for targets decreased (cold colors) or increased (warm colors) as a function of memory age (one hour to one month; voxel-wise p < 0.001, cluster-wise p < 0.05). Sections from anterior (upper left) to posterior (lower right). Higher F values indicate activity that more closely followed a power function. Red arrows and numbers highlight regions included in panel B. B. Patterns of brain activity (beta coefficients) in selected regions from the frontal and parietal lobes shown in A. Ang., Angular; ACC., Anterior Cingulate Cortex.; B., Bilateral; PCC., Posterior Cingulate Cortex; G., Gyrus.; IPL., Inferior Parietal Lobule; Med., Medial. MFG.; Middle Frontal Gyrus; R., Right. SPL.; Superior Parietal Lobule. Error bars show SEM.

Table 1.

Brain regions where retrieval-related activity was associated with the age of the memory

Brain Region	Vol. (mm3)	MNI Coordinate			B.A.	M.P.
		X	Y	Z
Increasing activity with memory age
B. Anterior Cingulate Ctx.	2862	1	19	39	32	↑
L. Sup. Frontal G.	243	−10	3	65	6	↑
R. Mid. Frontal G.	648	27	1	58	6	↑
L. Mid. Frontal G.	702	−25	−2	61	6	↑
L. Precentral G.	297	−27	−22	72	4	↑
R. Postcentral G.	297	55	−25	38	3	↑
R. Sup. Parietal Lob./Precuneus	1080	16	−66	60	7	↑
L. Sup. Parietal Lob./Precuneus	756	−20	−68	48	7	↑
R. Mid. Occipital G.+	270	33	−88	3	19	↑
Decreasing activity with memory age
B. Medial Frontal G.	972	6	49	−10	10,11	↓
L. Sup./Mid. Frontal G.	432	−29	27	47	8	↓
R. Sup./Mid. Frontal G.	972	32	20	52	8	↓
B. Anterior Cingulate Ctx.	243	1	−11	35	24	↓
L. Inf. Parietal Lob.	297	−59	−46	43	40	↓
B. Posterior Cingulate Ctx./ Precuneus	17955	−1	−53	34	23	↓
L. Supramarginal G./Sup. Temporal G.	810	−57	−61	29	40	↓
R. Inf./Sup. Parietal Lob./Angular G.	5994	45	−64	43	40	↓
L. Inf. Parietal Lob./Angular G.	1836	−43	−70	40	40	↓

Note. Activity in all clusters significantly changed across the four time periods according to a power function; voxel-wise threshold of p < 0.001, cluster-wise threshold of p < 0.05. For each monotonic pattern (M.P.) of activity across time periods [M.P., increasing (↑) or decreasing (↓)], clusters are listed from anterior to posterior based on the MNI coordinate of the center of mass. B.A., Brodmann area; B., Bilateral; Ctx., Cortex; G., Gyrus; Inf., Inferior; L., Left; Lob., Lobule; Mid., Middle; R., Right; Sup., Superior; Vol., Volume. Cross (+) denotes a cluster that was no longer significant after controlling for behavioral changes via the amplitude-modulation analysis.

Analysis of brain activity in hippocampal ROIs as a function of memory age

Analysis of brain activity in the bilateral hippocampal anatomical ROI did not reveal any significant changes in activity (beta coefficients) with memory age that followed a power function (LME Estimate: 3.93^e−04; SE: 1.22^e−02; df: 61.95; t = 0.03; p = 0.97) (Figure 4). The same was true for the beta coefficients from the amplitude-modulated analysis of activity as a function of memory age for the anatomical ROI (LME Estimate: −6.32^e−06; SE: 1.25^e−02; df: 61.98; t = 0.00; p = 0.99) and for the functional ROI (LME Estimate: −6.32^e−06; SE: 1.25^e−02; df: 61.98; t = 0.00; p = 0.99). Finally, a repeated measures ANOVA [Time period: hour, day, week, month] on the beta coefficients revealed no significant changes in activity across time periods for either the non-amplitude modulated (F(_3,60)=1.10, η² = 0.05, p = 0.36) or amplitude modulated (F(_3,60)=1.09, η² = 0.05, p = 0.36) analyses. The results were nearly identical for the left, right, anterior, and posterior anatomical hippocampal ROIs as well as the functional ROI (LME: Estimate:−4^e−0.3 – 0.01; SE: 0.008 – 0.02; df: 58.53 – 61.98; t = −0.23 – 0.94; p = 0.99 – 0.55; ANOVA: F(_3,60) = 0.54 – 1.26, η² = 0.03 – 0.06, p = 0.66 – 0.30). Thus, activity in the hippocampus did not significantly change as a function of memory age, regardless which method was used to carry out the analysis.

Figure 4.

Note. Pattern of brain activity (beta coefficients) in the bilateral anatomical hippocampal region of interest. No significant difference was detected across the timepoints (ANOVA: Memory Age)

Secondary analyses to clarify the role of brain regions in the memory-age network

Two secondary analyses were carried out identify if other factors could have mimicked or interfered with detection of the memory consolidation effects from the primary analysis. First, an amplitude-modulated analysis was carried out to examine if brain activity in the memory-age network might have been affected by the concomitant behavioral changes that were observed across the time periods (Figure 2; see Data Analysis, Secondary Analyses). All the regions in the memory-age network (see Table 1) were identified in the amplitude-modulated analysis, except for the cluster in the right middle occipital gyrus. No additional regions were identified. These findings indicate that almost all of the regions in the memory-age network were not affected by the concomitant changes in behavior across the time periods.

Brain activity associated with trial-by-trial variation in confidence level (Supplemental Table 1) or response time (Supplemental Table 2) were identified to examine the networks that support these behavioral measures. Brain regions where the association with behavior did or did not significantly change with memory age are reported separately. Of note, the brain regions which reflected response time variation were primarily in primary motor cortex, cerebellum, and PFC, whereas the brain regions which reflected confidence level variation were primarily in the MTL, PFC, and lateral parietal cortices.

Second, an analysis brain activity associated with successful encoding was carried out to identify if regions in the memory-age network might have been affected by re-encoding of the targets during the retrieval test (see Data Analysis, Secondary Analyses). A network of brain regions was identified where activity was significantly different for subsequently remembered versus forgotten foils that were implicitly encoded during the retrieval test (Supplemental Table 3). Importantly, this network of regions overlapped minimally with the memory-age network (Figure 5). This finding indicates that the regions in the memory-age network did not reflect re-encoding of the targets as a function of memory age. It is notable that the bilateral hippocampus and parahippocampal gyrus were identified by the encoding analysis (remembered > forgotten). This finding indicates that there was sufficient signal in the MTL to obtain this well-established effect (Miyamoto et al., 2014; Preston et al., 2010).

Figure 5.

Note. Minimal overlap of retrieval activity related to memory age and encoding-related activity. Brain regions where activity was related to memory retrieval of targets (blue) or incidental encoding (red) of foils. Retrieval regions are as shown in Figure 3. Activity in encoding regions significantly distinguished subsequently remembered foils versus subsequently forgotten foils based on the surprise post-scan recognition memory test (see Supplemental Table 3). Yellow voxels indicate areas of overlap between the two networks where green circles highlight the four small areas of overlap. A., Anterior; L, Left.

Brain regions where functional connectivity changed as a function of memory age

Functional connectivity with the vmPFC seed increased with memory age according to a power function in the left medial posterior parietal cortex (PPC) that included the precuneus and the posterior cingulate cortex (Figure 6, Table 2). This finding indicates that the influence of the vmPFC on medial PPC increased with memory age. A cluster in this region was also detected when the effects of concomitant behavioral changes were minimized (amplitude-modulated beta coefficients) (Table 2).

Figure 6.

Note. Ventromedial prefrontal cortex (vmPFC) and hippocampal functional connectivity. A. Sagittal section showing bilateral vmPFC seed region of interest (red). B. Coronal section showing bilateral hippocampal seed region of interest (red). C. Sagittal section showing a cluster in the posterior parietal cortex (PPC = posterior cingulate/precuneus) where functional connectivity with the vmPFC changed with memory age according to a power function (voxel-wise p < 0.001, cluster-wise p < 0.05). D. Coronal sections showing two clusters (see green circles: left parahippocampal cortex [PHC] and right fusiform gyrus; see Table 2) where functional connectivity significantly decreased with memory age (voxel-wise p < 0.001, cluster-wise p < 0.05). Exploratory analysis at a lower probability value (voxel-wise p < 0.01, cluster-wise p < 0.05) revealed that these regions were part of a larger network that included clusters in the PFC, insula, medial temporal lobe, lateral temporal lobe, posterior cingulate, parietal cortex, and occipital cortex (see Table 2). Higher F values indicate connectivity that more closely followed a power function. E. Pattern of functional connectivity (beta coefficients) between the vmPFC and the PPC for targets (black circles) or foils (white circle). F. Patterns of hippocampal functional connectivity for the regions circled in panel D for targets (black circles) and foils (white circles). L., Left; R., Right. Error bars show SEM.

Table 2.

Brain regions where retrieval-related functional connectivity was associated with the age of the memory

Brain Region	Vol. (mm3)	MNI Coordinate			B.A.	M.P.
		X	Y	Z
Connectivity with the vmPFC seed
L. Posterior Cingulate Ctx./Precuneus	1647	−16	−62	17	31	↑
AM Connectivity with the vmPFC seed
L. Posterior Cingulate Ctx./Precuneus	351	−19	−62	18	31	↑
Connectivity with the hippocampus seed
L. Parahippocampal Ctx./Cerebellum	270	−19	−41	−18	27	↓
AM Connectivity with the hippocampus seed
L. Parahippocampal Ctx./Cerebellum	405	−19	−41	−19	27	↓
R. Fusiform G.	243	42	−20	−22	20	↓
AM Connectivity with the hippocampus seed (exploratory analysis)*
R. Mid./Sup. Frontal G.	1107	31	17	51	8	↓
R. Parahippocampal G./Brain Stem	1161	10	−23	−34	28	↓
L. Insula/Transverse Temporal G.	918	−26	−23	21	13,41	↓
R. Hippocampus/Entorhinal Ctx./ Perirhinal Ctx./Fusiform G./ Inf./Mid. Temporal G./Caudate	4239	34	−27	−8	20,37	↓
L. Cerebellum	1674	−21	−39	−40	−	↓
L. Parahippocampal Ctx./Fusiform G./ B. Posterior Cingulate Ctx./ Retrosplenial Ctx./Cerebellum		−10	−44	−4	23,29, 30,31, 37	↓
B. Paracentral Lob./Precuneus/ L. Postcentral G.	1323	−2	−58	62	7	↓
R. Inf./Mid. Temporal G./Fusiform G./ Inf./Mid. Occipital G.	1674	46	−64	−8	20	↓

Note. Functional connectivity significantly changed across the four time periods according to a power function (voxel-wise threshold of p < 0.001 or *p < 0.01 [exploratory analysis], cluster-wise threshold of p < 0.05). AM (amplitude-modulated analysis) indicates that connectivity changed across time periods when the effect of concomitant changes in behavior were minimized. For each seed (ventromedial prefrontal cortex [vmPFC] or hippocampus), functional connectivity changed in a relatively monotonic pattern (M.P.) across time periods [M.P., increasing (↑) or decreasing (↓)].A.M., Amplitude-modulated; B.A., Brodmann area; B., Bilateral; Ctx., Cortex; Inf., Inferior; L., Left; Mid., Middle; R., Right; Sup., Superior; Vol., Volume.

Functional connectivity with the hippocampus seed decreased with memory age according to a power function in the left parahippocampal cortex (see Table 2). When the effects of behavioral changes were minimized (amplitude-modulated analysis), hippocampal connectivity continued to decrease with memory age in the left parahippocampal cortex and an additional cluster was detected in the right fusiform gyrus (Figure 6, green circles; Table 2). This finding indicates that the influence of the hippocampus on medial temporal and lateral temporal cortex decreased with memory age. To identify additional regions in the hippocampal network, we carried out an exploratory analysis of the amplitude-modulated beta coefficients at a lower probability threshold (p < 0.01, α = 0.05). A larger network was revealed that contained clusters in the PFC, insula, MTL, lateral temporal lobe, posterior cingulate cortex, parietal cortex, and occipital cortex, all of which exhibited decreased functional connectivity with the hippocampus with memory age (Figure 6, Table 2).

Connectivity associated with incidental encoding of the foils for the vmPFC and hippocampal seed regions was assessed to ascertain how connectivity during encoding was related to connectivity during retrieval. First, the retrieval-related connectivity map did not overlap with the encoding-related connectivity map for hippocampal (Supplemental Figure 1; Supplemental Table 4) or vmPFC connectivity (Supplemental Figure 2; Supplemental Table 5), suggesting that these maps do not overlap in the brain. Second, we asked if encoding-related connectivity could be considered the first step in the memory consolidation process. Specifically, we examined the amplitude of encoding-related connectivity in the regions identified by the primary functional connectivity analyses. For the PPC (identified by the vmPFC analysis) and the left parahippocampal cortex, and right fusiform gyrus (identified by the hippocampal analysis), encoding-related connectivity did not appear to reflect an early phase of the memory consolidation process. Rather, encoding-related connectivity in these regions was not part of the monotonic changes associated with memory retrieval (see Figure 6, white dots).

Discussion

We studied the consolidation of scenes as memories aged across one hour, one day, one week, and one month. Accuracy, discriminability, confidence ratings, and response time significantly changed with memory age. A broad network of cortical regions was identified, predominantly in the parietal lobe, where activity increased or decreased with memory age. Notably, the hippocampus was not part of this network. Changes in hippocampal activity with memory age were also not detected when using anatomically or functionally-defined ROIs. Functional connectivity between the vmPFC and the posterior parietal cortex increased with memory age. By contrast, functional connectivity between the hippocampus and left parahippocampal cortex decreased with memory age.

Secondary analyses were conducted to determine if the primary analyses, that examined retrieval-related activity and connectivity as a function of memory age, may have reflected other factors that changed with memory age, such as concomitant changes in behavioral measures and incidental encoding of targets. For brain activity, our primary findings were essentially unchanged after accounting for concomitant changes in confidence and response times (one out of 18 clusters was no longer significant, see Table 1). Importantly, the networks that support trial-by-trial variations in these behavioral measures correspond with networks known to support response time (Yarkoni et al., 2009) and confidence level (Kim, 2013; Moritz et al., 2006; Song et al., 2011; see Supplemental Tables 1, 2). There was also minimal overlap between the clusters that reflected encoding activity and the clusters that reflected retrieval activity as a function of memory age. Thus, our primary findings for brain activity did not appear to have been affected by these other factors.

For functional connectivity, the vmPFC findings were almost identical when concomitant changes in behavioral measures were reduced. By contrast, functional connectivity between the hippocampus and right lateral temporal cortex was detected only after controlling for the effects of behavior. Exploratory analysis of the hippocampal connectivity network revealed a larger network of regions including clusters in the PFC, cingulate gyrus, medial parietal lobe, and occipital lobe (as expected, all clusters exhibited decreases in connectivity with memory age). Functional connectivity of these seed regions (vmPFC and hippocampus) associated with successful incidental encoding of foils (remembered > forgotten) overlapped minimally with the retrieval-related connectivity reported in the primary analyses. Therefore, we believe that our primary analyses of brain connectivity also reflect changes in retrieval-related memory consolidation.

Patterns of memory consolidation in the cortex

Cortical Activity.

Systems consolidation theorizes that memories become hippocampus-independent as they are gradually reorganized in the neocortex (Alvarez & Squire, 1994; Marr, 1971; McClelland et al., 1995). Extending these ideas to brain activity measured with neuroimaging, the prediction is that cortical activity should increase with memory age. We identified an extensive network of cortical regions where activity increased with memory age (i.e., PFC, anterior cingulate cortex, parietal lobe, and occipital lobe). Increases in brain activity with memory age have also been observed in previous studies of consolidation over short time intervals. Although the specific brain regions do not overlap across previous studies (nor do they overlap substantially with the current study), increasing activity with memory age was most often observed in the frontal and parietal lobes (Davis et al., 2009; Gais et al., 2007; Suchan et al., 2008; Takashima et al., 2006; Yamashita et al., 2009), particularly in the precuneus (Sterpenich et al., 2009; Takashima et al., 2009). We also observed that precuneus activity (bilaterally) increased with memory age (see Figure 3A, clusters 4 and 5), though this activity was flanked on either side by clusters in posterior parietal cortex that decreased activity with memory age (see Figure 3A, clusters 8 and 9 in the inferior parietal lobules and cluster 7 in medial precuneus/posterior cingulate cortex). Thus, increasing and decreasing patterns of activation can occur adjacent to one another in the posterior parietal cortex.

We also observed relatively monotonic patterns of decreasing activity in the neocortex as memories aged. Regions that exhibited increasing activity with memory age were predominately located in the frontal lobe whereas regions that exhibited decreasing activity with memory age were predominately located in the parietal lobe. Although the finding that some cortical regions increased while others decreased with memory age is not predicted by systems consolidation, this pattern of findings has been observed in numerous studies of human neuroimaging of memory consolidation over short intervals (Bosshardt, Schmidt, et al., 2005; Harand et al., 2012; Takashima et al., 2017) and over longer intervals (e.g., years or decades; Douville et al., 2005; Gilboa et al., 2004; Haist, Bowden Gore, et al., 2001; Maguire et al., 2001; Niki & Luo, 2002; Smith & Squire, 2009).

Increasing cortical activity as a function of memory age across short intervals has also been observed in animal using immediate early gene expression or glucose metabolism as markers of brain activity. For example, activity in the anterior cingulate and/or PFC increased with memory age during context fear conditioning retrieval (Frankland et al., 2004; Wheeler et al., 2013), during retrieval of a learned spatial discrimination (Bontempi et al., 1999), and during retrieval of spatial memory (Teixeira et al., 2006). Similar to the findings for human cortical activity, some cortical regions exhibit decreases in activity with memory age (e.g., posterior cingulate cortex, Bontempi et al., 1999; Maviel et al., 2004).

Cortical Connectivity.

Systems consolidation theory also suggests that the complexity, distribution, and connections between cortical regions change as memories age (Alvarez & Squire, 1994; Frankland & Bontempi, 2005; Squire et al., 2015). We found that as memories aged, the influence of the vmPFC activity on medial PPC activity increased. Another study also found increasing connectivity between the PFC and lateral temporal cortex (Smith et al., 2010). Our findings demonstrate that connectivity between PFC and neocortex can also be observed in the parietal lobe. By contrast, correlations between PFC activity and hippocampal activity appear to decrease with memory age (Takashima et al., 2006; van Kesteren et al., 2010). We also observed decreasing connectivity between these regions, where the hippocampus exhibited a decreasing influence on several regions in PFC with memory age (see exploratory analysis of hippocampal connectivity; Table 2). Though these findings were identified at a less stringent threshold, they are nevertheless consistent with the idea that PFC-hippocampal activity decreases with memory age. Taken together, the increasing connectivity between the vmPFC and the neocortex and decreasing connectivity between the PFC and the hippocampus both support the prediction that long-term memory reorganizes in the neocortex after learning.

Findings from animal studies also reveal evidence cortico-cortical connectivity that increases as a function of memory age over this same time scale. In a massive study of context fear conditioning in mice, immediate early gene co-activation patterns were examined across 84 separate brain regions, including PFC, and compared when the memory was 1 day old or 36 days old. The findings unambiguously showed that neocortical regions were highly correlated with each other at the 36-day recall test versus the 1-day recall test (Wheeler et al., 2013). Similar findings were obtained for mPFC during retrieval of context-dependent, stimulus-stimulus associations (Morrissey et al., 2017; Takehara-Nishiuchi & McNaughton, 2008) and for posterior PFC-cortical connectivity increases across 1 month (Bontempi et al., 1999).

Patterns of memory consolidation in the hippocampus

Hippocampal Activity.

Studies of human neuroimaging and consolidation over short intervals (minutes to months) have found evidence that hippocampal activity decreases with memory age, supporting the prediction of systems consolidation theory (right hippocampus only, Bosshardt, Schmidt, et al., 2005; Dandolo & Schwabe, 2018; item memory, Du et al., 2019; Furman et al., 2012; Harand et al., 2012; Milton et al., 2011; Ritchey et al., 2015; Sekeres, Winocur, Moscovitch, et al., 2018; during delay period between visual cue and auditory stimulus, Smith et al., 2010; Sterpenich et al., 2009; Takashima et al., 2009; Takashima et al., 2006; Yamashita et al., 2009). However, other studies have found increases with memory age (Bosshardt, Degonda, et al., 2005; left hippocampus only, Bosshardt, Schmidt, et al., 2005; Gais et al., 2007; during visual cue, Smith et al., 2010), which are taken to support multiple-trace theory (MTT), wherein an increased number of memory traces in the hippocampus results in increasing hippocampal activity (Moscovitch & Nadel, 1998; Nadel & Moscovitch, 1997, 1998). We did not observe any of these relationships between hippocampal activity and memory age, even when using anatomical or functional hippocampal ROIs. Recent theoretical perspectives of memory consolidation make specific predictions about how hippocampal activity changes over time based on differences in vividness, details, and memory representations (Gilboa & Moscovitch, 2021; Sekeres, Winocur, & Moscovitch, 2018) or context (Yonelinas et al., 2019). We did not obtain measures of these components of memory, so future studies will need to test whether these components may illuminate the role of the hippocampus in retrieval as memories age.

Our review of the hippocampal activity findings in the 22 extant studies of consolidation over short intervals sheds light on which factors of these studies are associated with detecting a significant change in activity with memory age. We identified that studies with the longest time interval between recent and remote conditions were the most likely to detect decreasing hippocampal activity with memory age (see Supplemental Table 6). Specifically, 100% of the studies with a time interval greater than ~60 days found decreasing hippocampal activity with memory age (Furman et al., 2012; Harand et al., 2012; Milton et al., 2011; Sterpenich et al., 2009; Takashima et al., 2006; Yamashita et al., 2009), whereas only 60% of studies showed this pattern for time intervals between 27–42 days (Dandolo & Schwabe, 2018; item memory, Du et al., 2019; during delay between visual cue and auditory stimulus, Smith et al., 2010), and 36% of the studies showed this pattern for shorter time intervals (right hippocampus only, Bosshardt, Schmidt, et al., 2005; Ritchey et al., 2015; Sekeres, Winocur, Moscovitch, et al., 2018; Takashima et al., 2009). The increase in the probability of observing hippocampal activity that decreased with memory age was significantly related to these time intervals (Fischer’s exact test=6.5, p=0.035). For comparison, the findings for increasing activity in hippocampus over these time intervals were 0%, 40%, and 18%, respectively (Fischer’s exact test=2.6, p=0.211). The direction of this trend appears to be the opposite than would be predicted by MTT, where additional hippocampal traces would be more likely to occur with longer time intervals. Finally, studies that exhibited null results for hippocampal activity showed the same trend (0%, 40%, 55%, respectively; Fischer’s exact test=5.0, p=0.081). Therefore, it is possible that decreasing hippocampal activation is most readily detected when the interval between recent and remote memories exceeds 2 months, whereas the opposite trend occurred for increasing activity or null results. In light of this new analysis, our time interval (1 month) may not have been ideal for detecting differential hippocampal activity. Future studies seeking to test these theories by detecting differential activity would have better success if the time interval was 2 or more months.

Other factors also appear to affect the likelihood of detecting changes in hippocampal activity as a function of memory age. Out of the 9 studies that analyzed high-confidence hits or recollections from the Remember/Know procedure, 8 (~89%) found decreasing hippocampal activity as a function of memory age (Dandolo & Schwabe, 2018; Harand et al., 2012; Milton et al., 2011; Ritchey et al., 2015; Sterpenich et al., 2009; Takashima et al., 2009; Takashima et al., 2006; Yamashita et al., 2009) and the other study did not detect any changes in hippocampal activity as a function of memory age (Tompary & Davachi, 2017). However, when we analyzed only the high-confidence hits in the current study, we still failed to detect any difference in hippocampal activity with memory age. Other factors that we examined (type of memory [single-item vs. associative memory], choice of baseline, sample size, and laterality of hippocampal results]) did not appear to influence the probability of detecting differential hippocampal activity. Finally, it is worth emphasizing that poor signal in the inferior temporal lobes (due to their proximity to air/tissue and bone/tissue interfaces; Olman et al., 2009) can make it difficult to detect hippocampal activity. The current study did have sufficient signal in the hippocampus because we identified significant clusters in the hippocampus related to incidental encoding (i.e., subsequent hits > subsequent misses) and retrieval success (i.e., hits > misses). Thus, our lack of retrieval-related findings in the hippocampus as a function of memory age was not related to poor signal in the MTL.

Our examination of multiple time periods allowed for a more nuanced interpretation of our findings. Specifically, the majority of neuroimaging studies that investigated memory consolidation over short time intervals have examined only two time periods (a recent time period and a remote time period), and there was substantial variability in the age of memories assigned to the time periods (see Supplemental Table 6). While a significant decrease in activity across two time periods is typically taken as support for systems consolidation, a significant increase in activity across the time periods is typically taken as support for MTT. By contrast, when multiple time periods are examined, such as with the current study or six previous studies (Du et al., 2019; Furman et al., 2012; Gais et al., 2007; Stark & Squire, 2000; Suchan et al., 2008; Yamashita et al., 2009), one can determine if activity followed a pattern that is consistent or inconsistent with predictions. These types of studies provide stronger evidence because spurious factors that can cause differences in activity between two time periods can be distinguished from the factor of memory age (which increases in a monotonic fashion) across multiple time periods.

The pattern of hippocampal activity from the current study demonstrates this point (see Figure 4). For example, if we had only examined the two time periods of one hour and one day, we would have detected a decrease in hippocampal activity with memory age, supporting systems consolidation theory. If instead we had examined the two time periods of one day and one week, we would have detected an increase in hippocampal activity, supporting MTT. If we had examined one hour and one month, we would have found a null result for the effect of time period. By examining multiple time periods, we can report on the overall pattern of activity and our findings do not suffer from the difficulties of interpretation that may arise when only two time periods are examined. Thus, the pattern of hippocampal activity that we observed across time periods does not provide support for either theory.

Studies of hippocampal activity in animals across short time intervals using markers of brain activity such as glucose metabolism or immediate early gene activity also reveal evidence that hippocampal activity decreases with memory age (Bontempi et al., 1999; Frankland et al., 2004; Maviel et al., 2004; Wheeler et al., 2013). Nevertheless, there are reports of hippocampal activity that was similar or that increased for recent and remote retrieval of spatial memory (Makino et al., 2019; Teixeira et al., 2006), suggesting that hippocampal findings are also mixed for animals as well.

Hippocampal Connectivity.

Systems consolidation predicts that memory retrieval becomes hippocampus-independent as connections between the hippocampus and cortex are reorganized after learning (Squire & Alvarez, 1995). Although our findings for hippocampal activity do not align with the predictions of theories of memory consolidation, our findings for hippocampal connectivity do support the predictions of systems consolidation. Specifically, we found that functional connectivity between the hippocampus and MTL cortex (parahippocampal cortex) and lateral temporal lobe (fusiform gyrus) decreased with memory age, and the most dramatic change occurred across the first day. Another study of consolidation over short time intervals also found support for decreasing connectivity across 1 day between the hippocampus and parahippocampal cortex for word-scene pairs (Vilberg & Davachi, 2013). Our concordant findings demonstrate that decreasing connectivity between the hippocampus and MTL cortices can be detected within one day and can be obtained for both single item and associative memory for scenes. These findings for the parahippocampal cortex may reflect the importance of hippocampal connectivity for scene memory (i.e., parahippocampal place area; Epstein et al., 1999; Epstein & Kanwisher, 1998) as well as the importance of this connectivity for reorganization of memory for scenes over short intervals. Though we did not observe changes in parahippocampal cortex activity over short time intervals, and the results are mixed for this region across studies, decreasing activity with memory age is more consistently detected over long time intervals (years to decades) (Douville et al., 2005; Haist, Bowden Gore, et al., 2001; Smith & Squire, 2009; Woodard et al., 2010). Moreover, patients with lesions to the hippocampus and the surrounding MTL cortices (entorhinal, perirhinal, and parahippocampal cortex) exhibit greater temporally-graded retrograde amnesia than patients with lesions limited to the hippocampus (Bayley et al., 2006; Bright et al., 2006). Taken together, the parahippocampal gyrus appears to play an important role in long-term memory consolidation over both short and long time intervals. Finally, although memories are thought to be distributed in the cortex, damage to lateral temporal cortex (i.e., fusiform gyrus and inferior/middle/superior temporal gyrus) results in ungraded retrograde amnesia, suggesting that this region is important for retrieval of the content of memory (Bayley et al., 2005; Bright et al., 2006; Gilboa et al., 2005). Therefore, our finding that the influence of the hippocampus on lateral temporal cortex decreases with memory age suggests a role for the hippocampus in consolidation of the content of memory.

Exploratory analysis of hippocampal connectivity (at a lower probability threshold) revealed that the parahippocampal cortex and fusiform gyrus were part of a larger network that included the lateral PFC, insular cortex, MTL cortices, lateral temporal cortices, posterior cingulate cortex, parietal cortex, and occipital cortex. For all these regions, the influence of the hippocampus decreased with memory age. Our finding that the hippocampus exhibits a reduction in influence on PFC and lateral temporal cortices with memory age is consistent with prior studies of consolidation over short time intervals (Smith et al., 2010; Takashima et al., 2009; Takashima et al., 2006; van Kesteren et al., 2010). Hippocampal connectivity with the PFC, posterior cingulate, and precuneus also decreases with memory age over long intervals (Sheldon & Levine, 2013; Söderlund et al., 2012), suggesting that hippocampal connectivity with these regions is relevant for long-term memory consolidation over both short and long time intervals. It is important to emphasize that the broader hippocampus-cortical network was only detected when the effects of concomitant behavioral changes were minimized. Thus, hippocampal connectivity that tracks behavioral changes can interfere with the detection of connectivity that follows the age of memory. Future studies of systems consolidation will benefit from minimizing the effects of the behavioral changes that occur across the study-test interval.

Summary

In summary, we identified patterns of brain activation and connectivity that changed with the age of memory. Increases and decreases in activation were observed in the cortex, even though hippocampal activity did not exhibit any changes with memory age. Changes in cortico-cortical and hippocampal-cortical functional connectivity were consistent with systems consolidation theory. Measures of functional connectivity may be more useful for detecting the changing role of the hippocampus after learning.