Participants were randomly assigned to one of three conditions, each of which consisted of 10 training sessions, separated by at least one day over the course of approximately 4–8 weeks (4.12±1.87 weeks). Each training session lasted 2 hr. In the navigation training condition, we recruited a total of 29 participants, with 27 successfully completing 10 sessions of navigation training in a large city-scale virtual environment (‘Virtual Arida’, Figure 1b left). Briefly, participants were placed at one building and given the name of another building to find. Participants traveled to and from that pair of buildings until they were able to take a path not exceeding 120% of the shortest possible path. Upon reaching the criterion, they navigated between a new pair of landmarks. After one subsection of Virtual Arida was learned completely, participants were trained on integrating that subsection with previously learned subsections, with no more than one subsection learned in a day (see Methods). To quantify the training effect, we analyzed the average distance error traveled and normalized it by 120% of the length of the shortest possible path. While all 27 participants learned subsections A (although data was lost due to technical reasons for one participant), B, and A-B integration, 26 participants progressed to C, while 23 made it to A-B-C integration. 19 participants made it to subsection D, and 14 participants made it to A-B-C-D integration. Thus, to determine the training effect, we performed a repeated-measures ANOVA with subsection as the only factor and threshold-normalized distance error as the dependent measure.

We observed a significant main effect of subsection F(2, 42)=13.46, p<0.001, η²=0.39; removing sex and site from the model did not change the effect (Appendix 1). Participants improved over the course of the training, showing higher error initially for subsection A (M=79.09, SD = 33.22) compared to subsection B (M=65.52, SD = 26.89, Figure 1b right), which was significantly reduced for subsection C (M=47.44, SD = 31.22). This was also the case for subsections involving integration F(1,22) = 5.16, p=0.03, η²=0.19, although this finding did not survive inclusion of sex and site in the model (p=0.06). The distance error was reduced from subsections A-B (M=22.03, SD = 18.47) to A-B-C (M=15.17, SD = 14.47, Figure 1b right). In sum, participants improved both in learning new subsections of our large virtual training environment and in integrating multiple areas of the training environment together, suggesting that the navigation training resulted in significant reductions in path error over the course of the training.

In the verbal memory condition, we recruited a total of 32 participants, with 27 successfully completing 10 sessions of training. Briefly, participants viewed words for one second (one at a time), completed math problems as a distractor after all words were presented, and then recalled as many words as they could by typing them in any order they wished. They were trained on a time-based method of loci (‘temporal method of loci’) to improve their verbal memory (Bouffard et al., 2018). During the first five sessions, participants learned the same list of words, with the list growing in length after each successful trial (Same Word List, Figure 1a left). In the last five sessions, a new list of words was used for each trial (New Word List, Figure 1a left; see Methods). To quantify the training effect, we conducted a linear regression analysis on the maximum number of correctly recalled words for each day. The slopes of the regression lines for the first 5 days and the last 5 days were calculated separately (Figure 1a right), as the training paradigm differed between these two phases (see Methods). Both slopes were significantly greater than zero, indicating a significant improvement in memory performance over the training period: first 5 days, t(26) = 16.971, p<0.001; last 5 days, t(26) = 23.579, p<0.001. This effect remained significant after controlling for sex and site (Appendix 1), suggesting robust improvements in verbal memory performance as a result of the training.

In the active control condition, we recruited a total of 22 participants, with 21 successfully completing 10 sessions of watching videos. These videos included interviews, lectures, documentaries, and TV series on topics related to memory and navigation. After watching the videos for that session, participants responded to multiple-choice questions. We did not expect to find a training effect (i.e. improved performance on the questions over sessions), but did find that the percentage of correct responses each day was significantly higher than a 50% guess rate (Figure 1c right, ts(20) > 8.28, ps <0.001, mean = 88.58%, SD = 5.82%), indicating that participants were engaged with the video content.

We administered a set of cognitive tests related to navigation and verbal memory before and after training to track transfer from the training. This involved learning a large virtual environment not related to the training (i.e. Navigation Transfer task: Virtual Silcton Weisberg et al., 2014) and learning an untrained list of 12 words (i.e. Verbal Memory Transfer task) until all could be recalled perfectly (or after five trials). When we considered our pre-registered dependent measures specific to verbal memory (average number of words recalled, number of trials to criteria, and slope) and navigation (path error of Navigation Transfer task, pointing error of Navigation Pointing task, and map accuracy of Navigation Model Building task), we found some evidence of improvements in slope specific to the verbal memory training group and path error specific to the navigation training group. The other dependent measures, however, did not show any clear evidence of a training-specific benefit (see Appendix 1). We derived an exploratory measure that allowed us to compare between navigation and verbal memory, termed ‘learning rate’. The learning rate attempts to capture how quickly a participant might learn new information at post-test given that they were trained in improving their acquisition of this information. For example, participants who underwent navigation training should show a smaller difference in path distance on the first ten trials compared to the last ten at post-test compared to pre-test, reflective of faster learning due to the training (see Methods).

To quantify the learning rate for the Navigation Transfer task, a mixed-design ANOVA was conducted with condition (Navigation/Verbal Memory/Video Control) as a between-subjects factor and session (pre/post) as a within-subjects factor. This analysis revealed a significant main effect of session (F(1,70) = 37.52, p<0.001, η²=0.202) and a marginally significant main effect of condition (F(2, 70)=2.66, p=0.08, η²=0.028), with no significant interaction between condition and session (F(2, 70)=1.82, p=0.17, η²=0.020). Paired-sample t-tests indicated that all three groups demonstrated an increase in learning rate from pre to post, with the Navigation group showing the largest effect (Figure 2a; Navigation: t(26) = 4.43, p<0.001, Cohen’s d=1.32; Verbal Memory: t(26) = 2.31, p=0.03, Cohen’s d=0.596; Video Control: t(18) = 3.97, p<0.001, Cohen’s d=1.20). Results held when controlling for sex and site (see Appendix 1). The larger effect size for navigation suggests a greater benefit to the learning rate for the navigation condition, although this conclusion is tempered by the lack of a significant interaction effect.

To quantify the learning rate for the Verbal Memory Transfer task, we employed a mixed-design ANOVA, with condition (Navigation/Verbal Memory/Video Control) as a between-subjects factor and session (pre/post) as a within-subjects factor. We found a main effect of session (F(1, 71)=3.21, p=0.078, η²=0.014) and a significant interaction between condition and session (F(2, 71)=2.9, p=0.025, η²=0.034). Paired-sample t-tests indicated that only the Verbal Memory group showed a significant increase in learning rate from pre-test to post-test (Figure 2b; Verbal Memory: t(25) = 3.32, p=0.003, Cohen’s d=0.608; Navigation: t(26) = 0.488, p=0.63, Cohen’s d=0.113; Video Control: t(20) = –0.55, p=0.588, Cohen’s d=0.164). This effect remained significant after controlling for sex and site (see Appendix 1). These findings suggest that verbal memory training resulted in significant improvements from pre-test to post-test in the Verbal Memory Transfer task, specific to the Verbal Memory group.

As our task provided a separation of navigation and verbal memory skills, with some evidence of their independence in the transfer tests, we next tested the hypothesis that navigation training or verbal memory training (or both) would result in increases in hippocampal volume (as measured with T2 high-resolution imaging, see Methods). We conducted a mixed-design ANOVA with condition (Navigation/Verbal Memory/Video Control) as a between-subjects factor and session (pre/post) as a within-subjects factor. The analysis revealed no significant main effect of session (F(1, 71)=0.199, p=0.657, η²=0.0002, BF₁₀=0.151, moderate evidence against the inclusion of the main effect of session), no significant main effect of condition (F(2, 71)=0.113, p=0.893, η²=0.003, BF₁₀=0.241, moderate evidence against the inclusion of the main effect of condition), and no significant interaction between session and condition (F(2, 71)=1.224, p=0.300, η²=0.003, BF₁₀=0.054, strong evidence against the inclusion of the interaction effect). Paired t-tests further confirmed that none of the three training groups exhibited significant changes in total hippocampal volume from pre to post-test (Figure 3b, ps >0.265). This nonsignificant result persisted even after controlling for sex and site as covariates (ps >0.256).

We further investigated changes in hippocampal volume by conducting a mixed-design ANOVA, with a particular focus on lateralization. Specifically, we included regions of interest (ROI: left/right hippocampus) and session (pre/post) as a within-subjects factors, alongside condition (Navigation/Verbal Memory/Video Control) as a between-subjects factor. Consistent with the previous analysis, no significant main effects or interactions were observed (session: F(1, 71)=0.199, p=0.657, η²=0.0002, BF₁₀=0.131, moderate evidence against the inclusion of the main effect of session; condition: F(2, 71)=0.113, p=0.893, η²=0.003, BF₁₀=0.189, moderate evidence against the inclusion of the main effect of condition; ROI: F(1, 71)=0.208, p=0.650, η²=0.0001, BF₁₀=0.115, moderate evidence against the inclusion of the main effect of ROI; and session ×condition × ROI, F(2, 71)=0.072, p=0.930, η²<0.001, BF₁₀=0.0002, strong evidence against the inclusion of the interaction effect). Paired-sample t-tests confirmed that none of the three training groups exhibited significant changes in hippocampal volume from pre-test to post-test in either the left hippocampus (Figure 3b, ps >0.316) or the right hippocampus (ps >0.265). These findings remained consistent after controlling for sex and site as covariates (left hippocampus: ps >0.306; right hippocampus: ps >0.256).

Because some studies have shown long-axis specialization in correlations between hippocampal volume and behavior (Angeli et al., 2025; Brunec et al., 2018; Maguire et al., 2000), we considered whether anterior and posterior hippocampus might change differentially due to the training, which we analyzed based on the T1 whole brain structural image (see Methods). The previously described mixed-design ANOVA model was expanded to incorporate ROI (anterior and posterior hippocampus) as a within-subjects factor, thereby allowing for the examination of volume differences along the hippocampal anterior-posterior axis. Consistent with the above result, although we found a significant main effect of ROI (F(1,71) = 22.311, p<0.001, η²=0.052, BF₁₀ >100, strong evidence for the inclusion of the main effect of ROI), we found no significant main effect of session (F(1, 71)=2.914, p=0.092, η²<0.001, BF₁₀=0.439, weak evidence against the inclusion of the main effect of session), no significant main effect of condition (F(2, 71)=1.045, p=0.357, η²<0.022, BF₁₀=0.277, moderate evidence against the inclusion of the main effect of session), and no significant interaction effect between session ×condition × ROI (F(2, 71)=0.032, p=0.968, η²<0.001, BF₁₀=0.005, strong evidence against the inclusion of the main effect of condition). Simple main effects confirmed that none of the three training groups exhibited significant changes in hippocampal volume from pre-test to post-test neither in the anterior hippocampus (ps >0.272) nor the posterior hippocampus (ps >0.115). These findings remained consistent after controlling for sex and site as covariates (anterior hippocampus: ps >0.260; posterior hippocampus: ps >0.106).

Next, we divided the medial temporal lobe into six subregions: three hippocampal subfields (CA1, CA23DG, and SUB) and three adjacent subregions (PRC, ERC, and PHC). We then conducted a mixed-design ANOVA to examine changes in these subfields and subregions. Although results revealed a significant main effect of ROI (F(1.766, 125.390)=602.143, p<0.001, η²=0.895, Greenhouse-Geisser corrected; BF₁₀ >100, strong evidence for the inclusion of the main effect of ROI), we did not find any significant main effect of session (F(1,71) = 0.766, p=0.384, η²<0.001, BF₁₀=0.047, strong evidence against the inclusion of the main effect of session), condition (F(2,71) = 0.131, p=0.878, η²<0.001, BF₁₀=0.002, strong evidence against the inclusion of the main effect of session) and no significant interaction effect between session ×condition × ROI (F(3.21,110.806)=0.888, p=0.453, η²<0.001, Greenhouse-Geisser corrected; BF₁₀ <0.001, extremely strong evidence against the inclusion of the main effect of session). Paired-sample t-tests confirmed that none of the three training groups exhibited significant changes from pre to post-test in any of the six MTL subregions (Figure 3a, ps >0.588, FDR corrected). These findings remained nonsignificant after controlling for sex and site as covariates (ps >0.559, FDR corrected).

We also analyzed gray matter volume changes outside of the medial temporal lobe using FreeSurfer (see Methods) to determine if any cortical brain areas might have been affected by the training. We applied a vertex-wise analysis of cortical volume, again finding no significant differences across the entire cortex (see Methods). This finding was further validated using the Destrieux atlas (Destrieux et al., 2010), which included 74 cortical parcellations per hemisphere (148 ROIs in total). Paired-sample t-tests revealed that none of the ROIs exhibited significant volume changes from pre- to post-test in any of the three groups (ps >0.697, FDR-corrected). These findings suggest that training did not result in any measurable cortical volumetric changes.

For DWI, we separately evaluated free water (FW) and free water-corrected fractional anisotropy (fwcFA), as these measures capture different aspects of tissue microstructure. We also chose to analyze ROIs in gray matter (GM) and white matter (WM) separately, as these tissue types have dramatically different values for both FW and fwcFA, which involved eight GM ROIs and seven WM ROIs (see Methods). We conducted a mixed-design ANOVA with condition (Navigation/Verbal Memory/Video Control) as a between-subjects factor, with ROI and session (pre/post) as within-subjects factors. For the WM-fwcFA analysis, we observed a significant main effect of ROI (F(4.093, 233.279)=2340.878, p<0.001, Greenhouse-Geisser corrected, η²=0.628, BF₁₀ >100, extremely strong evidence for the inclusion of the main effect of ROI). Yet, we found no main effect of session (F(1,57) < 0.001, p=0.981, η²<0.001, BF₁₀=0.090, strong evidence against the inclusion of the main effect of session), no significant main effect of condition (F(2,57) = 0.025, p=0.975, η²<0.001, BF₁₀=0.011, strong evidence against including the main effect of condition), and no significant interaction between session ×condition × ROI (F(6.924, 197.328)=1.954, p=0.064, η²<0.001, BF₁₀ <0.001, strong evidence against including this interaction). Paired-sample t-tests confirmed that none of the three training groups exhibited significant changes from pre- to post-test in any of the seven ROIs (ps >0.083, FDR corrected). These findings remained nonsignificant after controlling for sex and site as covariates (ps >0.063, FDR corrected). Analyses involving WM-FW and GM regions (GM-fwcFA and GM-FW) were also not significant (see Appendix 1). These findings suggest that no DWI measures changed as a result of the training. Together, these findings suggest that neither navigation nor verbal memory training for 20 hr across 10 sessions was sufficient to induce changes in macroscopic hippocampal volume or microstructure of relevant gray matter regions and white matter tracts.

We then performed an exploratory analysis to determine whether individual differences in hippocampal volume independent of the training might predict how well they learned from pre-test to post-test based on the different trainings. We conducted analyses employing more stringent corrections for multiple comparisons to examine the association between hippocampal volume and the change in learning rate from pre- to post-test based on the Verbal Memory and Navigation Transfer tasks. This analysis addresses the idea of whether those with larger hippocampi might improve more from pre- to post-test based on the training. Because there were no pre-post differences in hippocampal volume, we averaged the pre- and post-test hippocampal volume. We consider other behavioral correlations with baseline hippocampal volume in the Appendix material; none of these were significant (see Appendix 1).

For the Navigation Transfer task, the learning rate did not correlate with hippocampal volume, hemispheric volume, anterior and posterior hippocampus, or any hippocampal subfields, which was also not significant when controlling for sex and site in all three groups (Appendix 1 and Appendix 2—table 5). For the Verbal Memory Transfer task, the changes in learning rate did not correlate with hippocampal volume (r(23) = 0.367, p=0.071). When accounting for sex and site as covariates, however, we found a significant positive correlation (r(23) = 0.605, p=0.006, FDR corrected). This effect was specific to the Verbal Memory group; no significant correlation was identified for either the Navigation group (r(25) = 0.028, p=0.889) or the Video Control group (r(19) = –0.206, p=0.371); these findings remained consistent when controlling for sex and site as covariates (Navigation: r(25) = 0.038, p=0.858; Video Control: r(19) = –0.123, p=0.858, FDR corrected).

We further examined the correlation between learning rate and hippocampal subfield volumes after correcting for multiple comparisons (see Methods), specifically focusing on CA1 and CA23DG. Only the CA23DG subfield showed a positive correlation with the change in learning rate from pre- to post-test in the Verbal Memory Group (r(23) = 0.532, p=0.036, FDR corrected, Figure 3c), and this correlation persisted after controlling for sex and site as covariates (r(23) = 0.493, p=0.017). No significant correlations were observed for the CA23DG subfield in the Navigation group (r(25) = –0.083, p=0.953, FDR corrected) or the Video Control group (r(19) = –0.12, p=0.953, FDR corrected), regardless of whether sex and site were included as covariates. For correlations by hemisphere, anterior/posterior hippocampus and other subfields, please see Appendix 1 and Appendix 2—table 4. These findings suggested that individuals in the Verbal Memory group with larger CA23DG volumes exhibited greater improvement in verbal memory performance from pre- to post-test.

Fisher’s z-tests revealed that the positive correlation in the Verbal Memory group was significantly stronger than that in the Navigation and Video Control group even after controlling for sex and site (Navigation: total hippocampus: Z=2.48, p=0.018; CA23DG: Z=1.972, p=0.036; Video Control group: total hippocampus: Z=2.596, p=0.015; CA23DG: Z=2.091, p=0.036; FDR corrected, one-side); there was no significant difference between the Navigation and Video Control group (total hippocampus: Z=–0.518, p=0.698; CA23DG: Z=–0.264, p=0.604; FDR corrected, one-sided). Together, these findings suggest that baseline hippocampal volume and CA23DG baseline subfield volume correlated with the change in learning rate on the Verbal Memory Transfer task following verbal memory training, but not after navigation training or video control. We highlight, however, that these analyses are exploratory and somewhat underpowered for between-subject comparisons.

Although we did not observe changes in hippocampal volume as a result of training, it is likely that other brain-related changes might be mediating the improvements specific to the navigation and verbal memory training. We investigated this issue by collecting fMRI data both pre- and post-training during a source memory encoding (Figure 4c) and retrieval task (Figure 5a). During the encoding phase, participants learned objects by placing them within either an autobiographical timeline (temporal) or a familiar spatial layout (spatial). During the retrieval phase, they saw the objects again and indicated which of the ‘sources’ (spatial or temporal) they had placed the object in. Because we trained the Verbal Memory group to use the temporal method of loci but trained the Navigation group to learn spatial locations, we reasoned that training-specific strategies might lead to context-specific changes (i.e. depending on whether the object was encoded in time or space) that would be reflected in network changes during the source memory tasks. We performed a univariate analysis across the whole brain and an informational connectivity analysis using 242 pre-defined brain regions involved in spatial processing and memory (see Methods, Figure 4b). Informational connectivity involved correlating multivariate patterns between trials within each ROI and then correlating again between all 242 ROIs (Figure 4a). Informational connectivity could reflect changes in network configurations during source memory encoding and retrieval based on skills gained from training.

We first considered both spatial and temporal source memory during the encoding phase, which could reveal how training affected the encoding of novel object-source pairings more broadly. For the univariate analyses, we found some changes from pre-test to post-test, but no interactions between session and training condition (see Appendix 1, Appendix 2—table 16 and Appendix 3—figures 6–8). When considering spatial or temporal encoding separately, we also found changes from pre-test to post-test, but no interaction between session and training group (see Appendix 1, Appendix 2—table 16 and Appendix 3—figure 6). All statistical analyses were corrected for multiple comparisons (see Methods).

Informational connectivity changes between pre-test and post-test were then compared across the three training groups. For the Verbal Memory group, we observed significantly increased informational connectivity at post-test compared to pre-test (i.e. post >pre), relative to the combined Navigation and Video Control groups (Figure 4). Specifically, we found enhanced connectivity between the left precuneus and left angular gyrus (t=3.94, p<0.05), the right lateral occipital cortex (LOC) and left intracalcarine (visual) cortex (t=3.63, p<0.05), and the right LOC and left superior parietal lobule (SPL) (t=4.16, p<0.05). In contrast, no significant decreases in informational connectivity (i.e. post <pre) were observed. Considering spatial and temporal encoding separately, for the Verbal Memory group during spatial encoding, we found enhanced informational connectivity from pre-test to post-test (i.e. post >pre) between the frontal cortex and visual cortex (see Appendix 1 and Figure 4d). For temporal encoding, increased connectivity from pre- to post-test (i.e. post >pre) was observed between the frontal cortex and temporal gyrus (Figure 4d). No significant decreases in informational connectivity from pre-test to post-test (i.e. post <pre) were observed in the Verbal Memory group relative to the combined Navigation and Video Control groups during either spatial or temporal encoding (Figure 4d).

Conversely, for the Navigation group compared to the combined Verbal Memory and Video Control groups, we observed significantly decreased informational connectivity at post-test compared to pre-test (i.e. post <pre). Specifically, reduced connectivity was observed between the left paracingulate gyrus and right middle frontal gyrus (MFG) (t=3.80, p<0.05), left superior frontal gyrus (SFG) and right supramarginal gyrus (SMG) (t=5.14, p<0.05), left frontal pole and right MFG (t=3.85, p<0.05), left dorsal LOC and right ventral LOC (t=3.7, p<0.05) and right MFG and left MFG (t=3.79, p<0.05). In contrast, no significant increases in connectivity were observed from pre- to post-test (i.e. post >pre) for the Navigation group relative to the combined Verbal Memory and Video Control groups (Figure 4d, bottom-left). When considering spatial and temporal encoding separately, during spatial encoding, reduced connectivity was observed from pre-test to post-test (i.e. post <pre) between the primary visual cortex and higher-order visual areas, while increased connectivity was identified from pre-test to post-test (i.e. post >pre) between the parietal and temporal cortices. For temporal encoding, the Navigation group exhibited significant decreases in connectivity from pre-test to post-test (i.e. post <pre) across multiple regions, including the frontal and parietal cortices (see Appendix 1), with no significant increases in connectivity from pre-test to post-test (i.e. post >pre). No significant changes in task-related informational connectivity to the hippocampus survived after multiple comparisons correction for either navigation or verbal memory training. Broadly, these findings suggest changes in informational connectivity specific to each training group (i.e. decreases in informational connectivity from pre-test to post-test in the Navigation group and increases in the Verbal Memory group), with no changes specific to the hippocampus.

For the univariate analyses for both spatial and temporal source memory retrieval, no significant changes were observed from pre-test to post-test, nor were there any interactions between session and training condition (see Appendix 2—table 16). When examining spatial and temporal retrieval separately, we similarly found no significant changes from pre-test to post-test and no interactions between session and training group (see Appendix 2—table 16).

We next examined informational connectivity changes from pre-test to post-test during the source memory retrieval stage (Figure 5a), finding a pattern somewhat opposite to that seen during encoding. In the Verbal Memory group, we observed significantly decreased informational connectivity in post-test compared to pre-test (i.e. post <pre) relative to the combined Navigation and Video Control groups. These reductions were identified between the right inferior frontal gyrus (IFG) and right frontal pole (t=4.01, p<0.05), left frontal pole and right SFG (t=4.48, p<0.05), right medial frontal gyrus and right posterior cingulate cortex (PCC) (t=4.42, p<0.05), right ventral LOC and left temporo-occipital fusiform cortex (TOFC) (t=4.30, p<0.05), and right lingual gyrus and left SPL (t=3.82, p<0.05). No significant increases in connectivity from pre-test to post-test (i.e. post >pre) were observed in the Verbal Memory group relative to the combined Navigation and Video Control groups (Figure 5b, top-left). Considering spatial and temporal source retrieval separately, spatial retrieval showed decreased connectivity from pre- to post-test (i.e. post <pre) between frontal regions (e.g. left and right superior frontal gyri) and between the right frontal pole and PCC, and increased connectivity from pre to post (i.e. post >pre) between temporal and parietal regions (e.g. left middle temporal gyrus [LMTG] and right SMG) and between the visual cortex and right SMG. For temporal retrieval, decreased connectivity from pre-test to post-test (i.e. post <pre) was prominent across multiple regions, including the frontal pole, angular gyrus, precuneus, and lateral occipital cortices. One increase specific to the Verbal Memory group was observed from pre-test to post-test (i.e. post >pre); however, in the left ventral lateral occipital cortex and left MFG (Figure 5b, see details in Appendix 1).

In the Navigation group during retrieval, we found significantly increased informational connectivity in the post-test compared to the pre-test (i.e. post >pre) relative to the combined Verbal Memory and Video Control groups. These increases were found between the right occipital pole and dorsal LOC (t=4.27, p<0.05), and the right SPL and right angular gyrus (AG) (t=4.09, p<0.05). Considering spatial and temporal retrieval separately, spatial retrieval showed enhanced connectivity from pre-test to post-test (i.e. post >pre) between regions such as the left frontal pole and the occipital and parahippocampal cortices. Temporal retrieval revealed increases from pre- to post-test (i.e. post >pre) between the precuneus and SFG, lingual gyrus and occipital gyrus, and SFG and MTG. No significant decreases in connectivity were observed from pre-test to post-test (i.e. post <pre) in the Navigation group relative to the combined Verbal Memory and Video Control groups (Figure 5b, bottom).

Again, no significant changes in task-related informational connectivity during source retrieval were present to the hippocampus when controlling for multiple comparisons. To summarize, verbal memory training resulted in decreased connectivity during retrieval across frontal, parietal, and occipital regions, especially during temporal retrieval, with limited increases. In contrast, navigation training consistently increased connectivity for both spatial and temporal retrieval, particularly between frontal, occipital, parietal, and temporal regions, with no significant decreases observed. These findings suggest that verbal memory primarily alters network connectivity during temporal contextual retrieval, whereas navigation training enhances connectivity more broadly across both contexts.

The above analysis focused on changes in informational connectivity between pre-test and post-test for individual connectivity edges between pairs of ROIs. Next, we examined changes in multivariate informational connectivity patterns, which we defined as the connectivity between a single ROI, and the remaining 241 ROIs. This analysis helped address network-wide changes in connectivity rather than those restricted to two different ROI pairs. Multivariate informational connectivity patterns were extracted separately for spatial and temporal informational connectivity matrices (ICMs), and the distance between the two patterns was computed using 1 r (Pearson’s, see Methods).

While no significant changes in pattern distance were observed during the encoding stage for any of the three conditions, a greater distance between spatial and temporal patterns in the left SFG network connectivity during source retrieval was found in the Verbal Memory group at post-test compared to pre-test (t(23) = –4.275, p=0.043, FDR corrected; Figure 6b). This effect was not observed in the Navigation and Video Control groups (Figure 6c). A similar result was obtained using the Spearman correlation to compute the distance, with a greater distance in the left SFG for the Verbal Memory group at post-test relative to pre-test (t(23) = –4.400, p=0.050, FDR corrected). No significant differences were found in the Navigation or Video Control groups. These findings also suggest some network-wide changes that are specific to verbal memory training.

Participants placed in the verbal memory condition completed free recall training on a desktop computer using a task built with Unity 3D. Similar to the verbal memory transfer task during pre/post-test, in this task participants were shown a list of words from the Toronto word pool one at a time, which they memorized using a trained strategy (the ‘temporal’ method of loci)(Bouffard et al., 2018). After being shown the list of words, they completed a distractor task by typing the numeric answer for five simple addition/subtraction questions. Then, they recalled the words by typing them into the provided space. The strategy that participants were given to memorize the words was a modified method of loci termed the ‘temporal method of loci’ (Bouffard et al., 2018). At the beginning of Day 1 of training, participants wrote 30 of their most memorable, vivid, and non-traumatic memories from their life thus far. They were asked to order these memories into a mental timeline to help them with the word recall task. Participants were asked to associate each word they were given with a memory from their mental timeline. Then, when they reached the recall portion of the task, participants were to use their mental timeline memories to help them recall the associated words in the current list. The number of words increased with each successful trial. If there were multiple unsuccessful trials, the number of words decreased. Once they were successful with this shorter list, the previous list was repeated. For example, if the list was ‘red’, ‘orange’, ‘yellow’, and the participant did not recall all three words, the next trial’s list would be ‘red’, ‘orange’. On successful completion, the next trial would be ‘red’, ‘orange’, ‘yellow’, ‘blue’.

 The pre-test and post-test navigation transfer task was completed on a laptop computer with an Xbox game controller using the ‘Virtual Silcton’ environment recreated in Unity 3D (Weisberg et al., 2014) (http://www.virtualsilcton.com/). Participants were instructed to freely navigate from their starting location to one of the eight possible destination buildings within the environment. If the participant walked for more than 60 s during a trial, a compass appeared at the bottom of the screen to point them to the target building. Over the trials, participants learned the spatial environment from the routes taken between the buildings. Participants completed trials until all routes were completed, or until 30 min elapsed. After the learning phase, participants completed an on-site pointing task, in which they were placed next to one of the eight buildings and had to point to each of the other seven in a random order. Next, they completed a model-building task, in which they dragged and dropped images of the eight buildings around an on-screen box. These tasks demonstrate configural knowledge of the environment.

Participants memorized and recalled a list of 12 words from the Toronto Word pool. The structure of the task was the same as that for the verbal memory training task, although it was programmed in PsychoPy. All participants had a maximum of five attempts to recall all 12 of the words correctly.

Prior to entering the scanner, participants were asked to generate a temporal and a spatial context to use during the source memory task. For their temporal context, they were asked to create a list of 10–15 salient memories of events from their life. We then asked them to write these down on a piece of paper and arranged them into a mental timeline of life events. For their spatial context, participants drew a blueprint of their living space (e.g. dorm room or apartment) on a second piece of paper. Participants were given time to review their contexts to ensure they knew them well enough prior to entering the scanner. They were then instructed on how they would use these contexts to recall objects. They were provided with an example of how to use each context and practiced verbally by using these contexts verbally.

 Participants underwent functional MRI scanning while performing the source memory task. This task was comprised of two phases, each with four runs. During the encoding phase (runs 1–4), participants were presented with images and instructed to encode them within either a temporal or spatial context, consistent with prior out-of-scanner practice. During the retrieval phase (runs 5–8), participants were presented with images and asked to make two judgments: first, whether the image was new, and second, if recognized, the context in which it had been previously encoded. All participants viewed the same images, but in counterbalanced orders.

To quantify the training effect, we analyzed the average distance error traveled, normalized by the threshold (120% of the length of the shortest possible path). We removed all trials that occurred after all unique pairs had been learned. Because we controlled the amount of time per task (2 hr per session), not all participants progressed to the same extent. Repeated-measures ANOVAs were conducted with subsection or integration as the within-subject factors and threshold-normalized distance error as the dependent measure.

To quantify the training effect, we performed linear regression analysis on the maximum number of correctly recalled words per day. Regression slopes were calculated separately for the first 5 and last 5 days (Figure 1a right), reflecting the distinct training paradigms implemented during these two phases (see Methods).

To assess navigation performance in the Navigation Transfer task, for each trial, we calculated distance error by subtracting the optimal distance from the participant’s actual navigated distance between their starting location and the target building. Since distance errors are inherently influenced by the distance between the starting location and the target building, we normalized these errors by dividing each distance error by the corresponding optimal distance. The task consisted of 20 trials, during which participants progressively learned the environment. As expected, normalized distance errors decreased over time, indicating improved navigation performance in later trials. To quantify this learning effect, we calculated a learning rate, defined as the negative of the difference between the normalized distance errors of the first 10 trials and the last 10 trials, divided by their sum (see Equation 1). A higher/positive learning rate suggests relatively stable and consistent performance across early and late trials, whereas a lower/negative learning rate reflects a greater difference in performance between early and late trials.(1)Learning Rate=−(Distance errorsoffirst10trials−Distanceerrorsoflast10trialsDistanceerrorsoffirst10trials+Distanceerrorsoflast10trials)\begin{document}$$\displaystyle Learning\ Rate=-\left (\frac{Distance\ errors\, of\, first\, 10\, trials\, -\, Distance\, errors\, of\, last\, 10\, trials}{Distance\, errors\, of\, first\, 10\, trials+\, Distance\, errors\, of\, last\, 10\, trials}\right)$$\end{document}

For the Verbal Memory Transfer task, we calculated a learning rate to quantify memory performance. To determine the learning rate, we first calculated the difference in the number of words correctly recalled between the first and last trials. This difference was then divided by the total number of trials to yield the learning rate (see Equation 2).(2) Learning Rate=Last trial−First trialnumberoftrials\begin{document}$$\displaystyle \ Learning\ Rate=\frac{Last\ trial-First\ trial}{\rm{number}\, \rm{of}\,\rm{trials}}$$\end{document}

To examine the relationship between learning rate of either task and brain structure (volume/DWI, see below), we collapsed the session factor by averaging values from pre- and post-training sessions for each ROI, given the absence of significant session effects or interactions between session and training condition. We conducted Pearson correlation analyses (parametric) for data that met the normality assumption based on the Shapiro-Wilk test. If normality was violated, Spearman correlation analysis (nonparametric) was used instead. Sex and site were included as covariates in the corresponding partial correlation analyses.

Scanning was performed using a 32-channel 3T Siemens ‘Skyra’ scanner at the University of Arizona and 32-channel 3T Siemens ‘Prisma’ scanner at the University of Florida. Visual stimuli were presented on a screen positioned behind the scanner and viewed by participants through a mirror attached to the head coil. Stimuli and responses were presented and collected using PsychoPy (https://www.psychopy.org) running on a Windows 10 laptop. High-resolution anatomical images of the hippocampus and surrounding cortex were acquired with a T2-weighted turbo-spin echo (TSE) anatomical sequence (FOV = 200 mm × 200 mm, matrix = 448 × 448, TR = 4200.0 ms, TE = 93.0 ms, flip angle = 139 degree, slice thickness = 1.9 mm, 28 slices, bandwidth=199 Hz/pixel). High-resolution structural images of the whole brain were obtained using a 3D, T1-weighted MPRAGE (1 mm³ isotropic) sequence (FOV = 256 mm, matrix = 256 × 256, slice thickness = 1 mm, TR = 2300 ms, TE = 2.41 ms, flip angle = 8 degree, bandwidth = 330 Hz/pixel). Functional images were acquired using a whole-brain echo planar imaging (EPI) sequence (TR = 1560 ms, TE = 30 ms, flip angle = 70 degree, field of view (FOV) = 220 mm, matrix = 88 × 88, slice thickness = 2.5 mm, slices = 48, bandwidth = 2030 Hz/pixel), involving a voxel resolution of 2.5 × 2.5 × 2.5 mm. Diffusion Weighted Imaging (DWI) data were acquired using two sequences with opposite phase-encoding directions to correct for distortion artifacts without signal loss. Acquisition parameters included: TR = 9200 ms, TE = 86 ms, FOV = 256 mm, 30 diffusion directions, 60 slices with a thickness of 2.0 mm, and a voxel size of 2.0 × 2.0 × 2.0 mm. The diffusion-weighted images were obtained using an echo plane sequence with the following parameters: number of b0 images = 1, b-value=1000 s/mm², number of directions = 30, TR = 9200 ms, TE = 86 ms, and voxel size: 2 × 2 × 2 mm. High-resolution resting state images were acquired using another whole-brain EPI sequence (TR = 3000ms, TE = 36 ms, flip angle = 90 degree, field of view (FOV) = 240 mm, matrix = 160 × 160, slice thickness = 2.5 mm, slices = 48, bandwidth = 802 Hz/pixel), involving a voxel resolution of 1.5 × 1.5 × 2.5 mm. Resting state data were not considered in this manuscript.

We performed pre-processing of DWI data using a customized pipeline that combines tools from the fMRIB Software Library 6.0 (Jenkinson et al., 2012), Advanced Normalization Tools (Tustison et al., 2021), and MRtrix3 (Tournier et al., 2019). First, we removed Gaussian noise present in the DWI data by fitting a Marchenko-Pastor distribution to the signal matrices to generate a threshold for PCA denoising (Cordero-Grande et al., 2019; Veraart et al., 2016a; Veraart et al., 2016b). We then removed Gibbs-ringing artifacts that can occur at tissue borders such as the outer surface of the brain and near ventricles (Tournier et al., 2019). Next, we generated brain masks for the DWI images using the dwi2mask function from MRtrix3, which uses information from both diffusion-weighted and non-diffusion weighted (b=0) volumes to generate an accurate brain mask. We then corrected for Eddy current and movement-related distortions using FSL Eddy (Andersson et al., 2016; Bastiani et al., 2019; Smith et al., 2004). Using the root mean square motion output provided by Eddy, we then applied a motion threshold, excluding participants with >2 mm absolute displacement or >0.5 mm relative displacement between diffusion weighted volumes. Application of this motion threshold at both timepoints resulted in 20 participants in the Verbal Memory group, 22 participants in the Navigation group, and 18 participants in the Video control group in final DWI analysis.

Consistent with prior work (Mitchell et al., 2022; Pasternak et al., 2009; Shin et al., 2025; Wilkes et al., 2024; Wilkes et al., 2023), we reconstructed the DWI data using a bi-tensor model, in which freely diffusing water is modeled by one tensor and anisotropic water diffusion is modeled with a separate tensor after removing the contribution from isotropic free water. In this model, the free water (FW) compartment of each voxel is interpreted primarily as originating from extracellular water diffusion, and the second tensor represents the tissue compartment after removing the contribution from FW. We performed whole brain estimation of FW and calculated free water corrected tensor metrics using custom MATLAB scripts (R2023a, The Mathworks, Natick, MA, USA). Briefly, we calculated FW from single shell diffusion data based on minimization of a variational regularization framework outlined in Pasternak et al., 2009. Initialization of the free water estimate in this pipeline used mean diffusivity (MD) maps calculated from a single tensor fit. Thus, prior to FW estimation, we performed conventional single tensor reconstruction using FSL’s DTIFIT. Next, conventional MD maps were used for initialization of the bi-tensor reconstruction in MATLAB. Voxels with MD values greater than 0.8 x d\begin{document}$d$\end{document} (i.e. d\begin{document}$d$\end{document} is constant diffusivity) were assumed to be comprised of CSF and omitted from the fitting process. Fitting a bi-tensor model with DWI data from a single non-zero diffusion weight (i.e. b-value) has been described as an ill-posed problem with nearly infinite solutions, thus we employed sensible biological constraints to the minimization process, as performed in Pasternak et al., 2009. We set the reference MDt\begin{document}$MD_{t}$\end{document} at 0.6 μm2/ms\begin{document}$\mu m^{2}/ms$\end{document}. and diffusivities were limited to λmax\begin{document}$\lambda _{max}$\end{document} = 2.5 μm2/ms\begin{document}$\mu m^{2}/ms$\end{document} and λmin\begin{document}$\lambda _{min}$\end{document} = 0.1 μm2/ms\begin{document}$\mu m^{2}/ms$\end{document}. We also assumed isotropic water diffusion at 37 ° C, corresponding to human body temperature, as a constant (d\begin{document}$d$\end{document}=3.0 x 10^–3). After initialization, 100 iterations were used to refine the FW estimates corresponding to free water corrected metrics. We used an automated quality assurance procedure and visually inspected output images (e.g. FW and fwcFA) to confirm appropriate data quality.

We used ANTs to perform nonlinear registration to MNI standard space by warping each participant’s FA image to the HCP 1065 template (Yeh, 2022). We used participants’ uncorrected FA for the registration process because the HCP 1065 template was created with FA images from single tensor reconstruction, not fwcFA images. We then applied the same transformation matrix to align FW and fwcFA images from the same participant into MNI space. Following registration, we extracted FW and fwcFA values from ROIs in MNI space. We separately evaluated ROIs from gray matter (GM) regions and white matter (WM) tracts, selected based on prior literature and relevance to navigation. Eight bilateral GM ROIs from the Mayo Clinic Adult Lifespan Template (Schwarz et al., 2017) were used: hippocampus, parahippocampal gyrus, entorhinal cortex, inferior temporal cortex, cuneus, superior parietal cortex, retrosplenial cortex, and caudate nucleus (Ekstrom et al., 2017). Seven bilateral WM ROIs from the Johns Hopkins University (JHU) white-matter tractography atlas were used: three subregions of the corpus callosum (body, genu, and splenium), fornix, posterior thalamic radiation, cingulum (hippocampus), and the fornix cres / stria terminalis (these two small tracts are not able to be differentiated at 2 mm voxel resolution).

 Similar to previous work (Wilkes et al., 2024), we performed separate statistical analyses of ROIs from GM and WM regions, as these tissue types have dramatically different values for DWI metrics. We also separately evaluated FW and fwcFA, as these metrics provide separate but complementary information about extracellular and intracellular contributions to tissue microstructure. For each of these four combinations (GM-FW, GM-fwcFA, WM-FW, WM-fwcFA), we performed mix-designed ANOVAs with condition (Navigation/Verbal Memory/Video Control) as a between-subjects factor and ROI and session (pre/post) as within-subjects factors with corresponding post-hoc comparisons. We also performed correlations (Pearson or Spearman, as appropriate) to evaluate the relationship between DWI metrics and performance on behavioral measures. Correlations were corrected for multiple comparisons using the false discovery rate (FDR) method (Benjamini and Hochberg, 1995), which was applied across p-values for each ROI, performed separately for FW and fwcFA analyses.

Automatic hippocampal subfield segmentation software (ASHS) (Yushkevich et al., 2015a; Yushkevich et al., 2015b) was used to segment the subfields of the MTL based on each participant’s high-resolution T2-weighted MRI image. We used ASHS with the ASHS-Princeton-1.0.0-Young-Adult (Yushkevich et al., 2015a). The MTL was segmented into CA1, CA2/3, DG, and subiculum (SUB), as well as the perirhinal cortex (PRC), entorhinal cortex (ERC), and parahippocampal cortex (PHC). We combined the CA2/3 and DG subfields as finer distinctions could not be made at the acquired resolution (Zeineh et al., 2001). Each participant’s subfield segmentations were manually inspected to ensure accuracy of the segmentation protocol. The hippocampus was further subdivided into anterior (head) and posterior (body +tail) regions along its longitudinal axis using the T1-weighted MPRAGE sequence and FreeSurfer 7.4.1 software (https://surfer.nmr.mgh.harvard.edu/). To ensure reliable volume estimates, images were processed using FreeSurfer’s longitudinal pipeline (Reuter et al., 2012). This method creates an unbiased within-subject template using a robust, inverse-consistent registration approach (Reuter et al., 2010). Visual quality control was conducted by three trained raters who inspected the skull stripping, surface reconstruction, and segmentation accuracy at both the within-subject template and individual timepoints. Manual edits were primarily performed on the within-subject template to correct segmentation errors—particularly in challenging regions such as the hippocampus—since edits to the template automatically propagate to all associated timepoints. Raters followed standardized FreeSurfer longitudinal editing guidelines to ensure consistent correction across subjects and time points. Discrepancies between raters were resolved via consensus discussions. This approach enhanced segmentation accuracy and consistency across longitudinal scans, thereby improving the reliability of volumetric and morphometric analyses. These quality control procedures ensured that anatomical registration was accurate and that atlas-based ROI extractions were not confounded by segmentation or alignment errors. To account for individual differences in brain size, we normalized the volume of each MTL subregion by dividing each volume by intracranial volume (ICV) and scaling by a factor of 100.

We conducted whole-brain structural analyses using FreeSurfer (version 7.4.1; https://surfer.nmr.mgh.harvard.edu). T1-weighted anatomical images were processed with the longitudinal processing pipeline (Reuter et al., 2012).

Vertex-wise analyses of cortical volume were conducted using FreeSurfer’s general linear modeling tool, mri_glmfit. Group-level statistical comparisons were corrected for multiple comparisons using mri_glmfit-sim, which implements cluster-wise correction based on Monte Carlo simulations. A vertex-wise threshold of Z>3.0 (corresponding to p<0.001, two-sided) was used to identify both positive and negative effects. Clusters were considered significant if they survived a cluster-wise corrected p-value of less than 0.05.

In addition to vertex-wise analyses, cortical parcellation was performed using the Destrieux atlas (Destrieux et al., 2010) (aparc.a2009s), which delineates 74 cortical regions per hemisphere, resulting in a total of 148 cortical regions of interest (ROIs). To account for inter-individual variability in brain size, each ROI volume was normalized by the estimated intracranial volume (ICV) and scaled by a factor of 100. Longitudinal statistical comparisons were conducted using paired-sample t-tests. To control for multiple comparisons across ROIs, FDR correction was applied at q<0.05.

To identify brain regions associated with memory and spatial navigation, we conducted meta-analyses using Neurosynth (https://neurosynth.org/). We first defined 11 maps based on Neurosynth association tests, using the following 11 key terms: ‘memory’, ‘episodic memory’, ‘autobiographical memory’, ‘recognition memory’, ‘subsequent memory’, ‘memory encoding’, ‘memory retrieval’, ‘memory performance’, ‘memory processes’, ‘memory tasks’, and ‘memory test’. These maps were then combined to create a single, comprehensive memory-related brain mask. Then, following the same steps, we created a single, comprehensive spatial navigation-related mask using six key terms: ‘navigation’, ‘spatial’, ‘spatial information’, ‘spatial temporal’, ‘time task’, and ‘visual spatial’. The memory-spatial map was then created by summing up the memory-related and navigation-related maps. Finally, this combined map was overlapped with the 400 ROIs from the ‘Schaefer 2018 parcellation’ to yield 242 ROIs, each containing a minimum of 30 voxels. Prior to ROI extraction, all registration steps—from individual subject space to MNI space—were visually inspected for each participant to ensure accurate alignment between the functional data and the Schaefer atlas.

We performed fMRI data preprocessing using FEAT (FMRI Expert Analysis Tool), version 6.00, implemented in FSL (http://www.fmrib.ox.ac.uk/fsl). The EPI images underwent motion correction, slice-timing correction, and temporal filtering with a nonlinear high-pass filter (100  s cutoff). Six motion parameters were included as confounding regressors in the model. Additionally, outlier timepoints which were identified using the FSL’s motion outlier detection tool (framewise displacement [FD]>0.9 mm) were incorporated as additional confounds in the first-level general linear model (GLM) analysis. If more than 20% of the volumes in a run had an FD exceeding 0.9 mm, or if the absolute head movement exceeded half the voxel size (i.e. 1.25 mm), the entire run was excluded from further analyses. For single-trial estimation, no spatial smoothing was applied. All functional images were linearly registered to the middle image of the first run, and all analyses were conducted in MNI standard space.

We examined encoding and retrieval-related neural activity using the general linear model (GLM) within the FILM module of FSL. During the encoding stage, the GLM included five regressors: (1) the encoding stage of all pictures (i.e. the 6 s of the trial in which the participants were presented with an image and asked to encode that image using a temporal or spatial context); (2) rating period (i.e. reaction time, RT) when the participant made a rating judgment of all trials; (3) the remaining time period after the participant made a rating judgment of all trials (i.e. 6 s minus RT) while the stimulus stayed on the screen; (4) The duration of inter-trial interval (ITI) fixations; (5) all the instruction period within the same run. The ‘X’ or ‘O’ judgement baseline period was not coded and thus was treated as an implicit baseline. Events were modeled at the time of the stimulus onset and convolved with canonical hemodynamic response function (double gamma function). The encoding phase consisted of four runs, equally divided between spatial and temporal encoding runs. For the retrieval phase, remembered and forgotten pictures were separately modeled for both item memory retrieval and source memory retrieval. Source memory retrieval was further categorized into spatial and temporal retrieval trials. The source memory retrieval effect was defined as the difference in activity between correctly retrieved contextual pictures and the implicit baseline.

A second-level analysis was conducted to compute cross-run averages for spatial encoding effects, temporal encoding effects, combined encoding effects (spatial and temporal concatenated), spatial source memory retrieval effects, temporal source, and combined source memory retrieval effects (spatial and temporal concatenated) using a fixed-effects model. These contrasts were then used in a group analysis with a random-effects model using full FMRIB’s Local Analysis of Mixed Effect 1 with automatic outlier detection (Beckmann et al., 2003; Woolrich, 2008). Unless otherwise noted, group images were thresholded using cluster detection statistics, with a height threshold of z>3.1 and a cluster probability of p<0.05, corrected for whole-brain multiple comparisons using Gaussian Random Field Theory. To specifically examine univariate activation within the hippocampus, a key region of interest in our experiment, we conducted small volume correction using the hippocampal mask derived from the Harvard-Oxford subcortical atlas in MNI152 space.

General linear models (GLMs) were conducted separately to estimate the activation patterns for each of 80 encoding trials and 100 retrieval trials. In each single-trial model, a Least Square–Separate (LS-S) approach was used, in which the trial of interest was modeled as one regressor, with all other trials modeled as a separate regressor (Mumford et al., 2012). Specifically, during the encoding stage, each single-trial GLM included six regressors: (1) the trial of interest (i.e. the 6 s of the trial where the participants were presented with an image and asked to encode that image using a temporal or spatial context); (2) all other remaining trials within the same run; (3) rating period (i.e. reaction time, RT) when the participant made a rating judgment of all trials; (4) the remaining time period after the participant made a rating judgment of all trials (i.e. 6 s minus RT) while the stimulus stayed on the screen; (5) The duration of inter-trial interval (ITI) fixations; (6) the instruction period within the same run. The ‘X’ or ‘O’ judgement period was not coded and thus was treated as an implicit baseline.

Similarly, for the retrieval stage, each single-trial GLM included six regressors: (1) the trial of interest (i.e. the RT of the trial in which participants were asked to retrieve source information for the image); (2) all other remaining source retrieval trials within the same run; (3) the remaining time period after the participant made a source judgment (i.e. 6 s minus RT) while the stimulus remained on the screen; (4) all item retrieval periods in which participants judged whether they could recognize the item from the encoding phase by pressing ‘Yes’ or ‘No’ within the same run; (5) the remaining time period after the participant made a Yes/No judgment (i.e. 6 s minus RT) while the stimulus stayed on the screen; and (6) all instruction periods within the same run. The ITI fixation period was not modeled and was treated as an implicit baseline. Each event was modeled at the time of stimulus onset and convolved with a canonical hemodynamic response function (double gamma). To control the effects of head motion, six motion parameters were included in the GLM model as a covariate, as well as a regressor for each TR that was flagged as having greater framewise displacement (FD) than 0.9 during preprocessing. The t-map for each trial was used for multivariate pattern similarity analysis to increase the reliability by normalizing for noise (Walther et al., 2016).

Multi-voxel pattern similarity (MPS) was used to measure the similarity of activation patterns by calculating the correlation between trials that were either encoded or correctly retrieved, within each ROI (Kriegeskorte et al., 2006). Following the approach of Power (Power et al., 2014), volumes with a framewise displacement greater than 0.9 mm were censored, and trials containing any censored frames within the duration of the modeled GLM response were excluded. Specifically, within-context pattern similarity (PS) during encoding was calculated using pairwise Pearson correlation coefficients between trials within the same context (i.e. spatial or temporal). Within-spatial context was calculated using only spatial trials, while within temporal context was calculated using only temporal trials. For source memory retrieval, we performed analogous calculations for within-context PS, separately measuring within-spatial context PS and within-temporal context PS, using correct retrieval trials. All MPS analyses were performed across trials from different runs to avoid temporal autocorrelations that could otherwise inflate or bias results. The resulting correlation coefficients were then transformed into Fisher’s z-scores.

We performed informational connectivity among the 242 predefined ROIs by calculating pairwise correlations between representational similarity matrices (RSMs) (Aly and Turk-Browne, 2016; Gao et al., 2022; Kriegeskorte et al., 2008; Zheng et al., 2023). This was done by correlating the within-context PS for each ROI with another ROI. In contrast to traditional functional connectivity (FC) analyses, which quantify temporal correlations of BOLD signal fluctuations between brain regions, the informational connectivity approach examines the similarity of information coding patterns across regions (Aly and Turk-Browne, 2016; Gao et al., 2022; Zheng et al., 2023). While FC analyses rely on averaged BOLD signals within regions, representational similarity analysis (RSA) captures the fine-grained spatial patterns of neural responses to different experimental conditions (Kriegeskorte et al., 2008). Specifically, RSA quantifies how similarly or differently a brain region responds to different stimuli or conditions by comparing multivariate activity patterns, creating a RSM that characterizes each region’s representational structure (Kriegeskorte et al., 2008). By correlating these RSMs between regions, we can assess whether different brain areas encode information in similar ways, providing insights into information sharing and processing that complement traditional FC measures.

Three types of within-context RSMs were calculated: one based on all trials, one based on spatial trials, and one based on temporal trials (Figure 4a top). All Pearson correlation coefficients (r values) were Fisher-Z transformed prior to statistical analysis. Significance thresholds for informational connectivity were determined using 10,000 permutation simulations. The reported results have been corrected for multiple comparisons using the FDR method, with a q-value threshold of less than 0.05.

To assess changes in multivariate informational connectivity patterns between pre and post, distances were computed for each ROI by comparing spatial and temporal informational connectivity patterns before and after training. Specifically, for each ROI, the informational connectivity with the remaining 241 ROIs was extracted separately from the spatial and temporal informational connectivity matrices (ICM). The distance between the two connectivity patterns was calculated using 1 minus the Pearson or Spearman correlation coefficient (Figure 6a). Paired t-tests were conducted to evaluate statistical differences between pre and post, with FDR correction applied across the 242 ROIs at a q-value threshold of <0.05.