Cortico-hippocampal network connections support the multidimensional quality of episodic memory

Episodic memories reflect a bound representation of multimodal features that can be reinstated with varying precision. Yet little is known about how brain networks involved in memory, including the hippocampus and posterior-medial (PM) and anterior-temporal (AT) systems, interact to support the quality and content of recollection. Participants learned color, spatial, and emotion associations of objects, later reconstructing the visual features using a continuous color spectrum and 360-degree panorama scenes. Behaviorally, dependencies in memory were observed for the gist but not precision of event associations. Supporting this integration, hippocampus, AT, and PM regions showed increased connectivity and reduced modularity during retrieval compared to encoding. These inter-network connections tracked a multidimensional, objective measure of memory quality. Moreover, distinct patterns of connectivity tracked item color and spatial memory precision. These findings demonstrate how hippocampal-cortical connections reconfigure during episodic retrieval, and how such dynamic interactions might flexibly support the multidimensional quality of remembered events.


INTRODUCTION
Memories for past events are highly complex, allowing us to travel back in time and subjectively reexperience episodes in our lives. These events are not stored and played back to us as we experienced them; rather, they are reconstructed in a hierarchical manner. Episodic reconstruction is thought to be facilitated by hippocampalneocortical processes that rebuild the rich content and quality of past events within a spatiotemporal framework (Barry & Maguire, 2018; Ranganath, 2010; Ritchey, Libby, & Ranganath, 2015; Robin, 2018 and integrate them with prior knowledge (Morton, Sherrill, & Preston, 2017) . In turn, this adaptive, reconstructive process can lead to forgetting of specific event features and variability in the precision which different features are remembered (Schacter, Guerin, & St Jacques, 2011) .
Previous research has found widespread increases in cortical and subcortical brain activity when people successfully remember rather than forget events (Rugg & Vilberg, 2013) . Beyond changes in activity, largescale brain networks increase their communication strength during episodic retrieval tasks (Fornito, Harrison, Zalesky, & Simons, 2012; Robin et al., 2015; Westphal, Wang, & Rissman, 2017 , where functional connectivity, particularly of the hippocampus, is increased when events are remembered compared to forgotten (Geib, Stanley, Dennis, Woldorff, & Cabeza, 2017; King, de Chastelaine, Elward, Wang, & Rugg, 2015; Schedlbauer, Copara, Watrous, & Ekstrom, 2014; St Jacques, Kragel, & Rubin, 2011 . Such neural changes are validated by behavioral evidence showing that event features are dependent on one another in memory, emphasizing that remembering involves the binding of distinct elements into a single, coherent event representation (Horner & Burgess, 2013 . This binding process is widely thought to be facilitated by the hippocampus (Barry & Maguire, 2018; Horner, Bisby, Bush, Lin, & Burgess, 2015; Moscovitch, Cabeza, Winocur, & Nadel, 2016; Ritchey, Libby, et al., 2015 . Therefore, episodic retrieval is likely dependent on the coordination of memory 'hubs' such as the hippocampus with neocortical regions to reconstruct and integrate the diverse components of memory representations. Despite this research, little is known about how changes in hippocampalcortical communication flexibly support the multidimensional quality of remembered events. Distinct cortical areas support the different building blocks of episodic memory: for instance, parahippocampal cortex (PHC) is thought to provide the hippocampus with spatial context information, whereas content and objective precision of retrieved events. To this end, we tested participants on a memory reconstruction task to obtain continuous measures of different episodic memory features (Brady, Konkle, Gill, Oliva, & Alvarez, 2013; Harlow & Yonelinas, 2016; Nilakantan, Bridge, Gagnon, VanHaerents, & Voss, 2017; Richter et al., 2016 . Participants learned a series of objects, each with a color, scene location, and emotion association, and then reconstructed the visual appearance of the objects later on. Here, they selected a color from a continuous spectrum and moved around 360° panorama scenes to place the object in its original location, providing a sensitive, objective, and naturalistic way of assessing memory (cf. Serino & Repetto, 2018) . We predicted that PM and AT systems would show a distinct network structure during encoding, but that, crucially, these networks would become more integrated during episodic retrieval. Moreover, we expected that increased internetwork and hippocampal connectivity would dynamically track binding and the composite quality of features within memory. In line with the representational organization of the PMAT framework, we finally predicted that functional connectivity of PM and AT systems would track memory precision for spatial context and item information, respectively.

RESULTS
Participants completed an episodic memory task in which they learned 3 features associated with trialunique objects: a color from a continuous spectrum, a location within a panorama scene, and an emotionally negative or neutral sound ( Figure 1A). In a subsequent test, participants were first cued to covertly retrieve as much information about each object as possible, and then they dynamically reconstructed each object's color and scene location ( Figure 1B), providing continuous measures of memory error in degrees (remembered feature value encoded feature value). Using these finegrained memory measures, we test how the content and fidelity of information is bound into a single memory representation, and if these memory processes are supported by flexible engagement of the PM and AT corticohippocampal networks (Ritchey, Libby, et al., 2015) .

Episodic features are recollected with varying precision
We first evaluated behavioral performance to quantify memory variability, both in terms of the probability of successful retrieval and precision of each reinstated feature. The proportion of 'correct' responses (memory success) was calculated for each of the object features item color, spatial context, and emotion association and the precision of correct recollection was additionally estimated for color and spatial features. Here memory performance was evaluated by fitting a mixture model (Bays, Catalao, & Husain, 2009; Zhang & Luck, 2008 to each subject's errors ( Figure 2A; see Methods). Participants remembered the features well above chance on average (Table 1), and the proportion of correct responses did not differ between color and scene features (t(27) = 1.09, p = .29). Participants varied in the precision (k) with which they could remember these visual details, but were more precise when remembering the object's spatial location compared to color (t(27) = 6.48, p < .001).

The gist but not precision of episodic features is bound in memory
Based on the hypothesis that interactions between hippocampus and the PM and AT systems support the integration of recollected episodic information, we sought to test if measures of memory success and precision were dependent across features (see Methods). We expected that successful retrieval of one feature would promote memory for the others (Horner & Burgess, 2013 . We additionally asked whether successful retrieval further influences the precision with which visual information is remembered, and is the precision of different features in memory related? All feature pairs showed significant memory dependency for successful vs. unsuccessful retrieval ( Figure 2B upper panel; ts(27) > 7.30, ps < .001), so that retrieval of one feature was likely to lead to successful retrieval of the others. However, successful recollection of color and scene information did not significantly benefit the precision with which the other feature was recalled (ts < 1.85, p s > .07). Color and scene memory precision were also unrelated (t = 0.32, p = .75) ( Figure 2B lower panel). Therefore, integration of episodic information into a coherent memory trace likely involves the binding of gistlike information about distinct features, whereas the fidelity of each feature in memory appears to be somewhat independent of this process. correlations between each ROI ( Figure 3A) times series were computed across encoding and remember task events after first regressing out all trial and memoryrelated activity and nuisance variables such as motion (see Methods). Thus, connectivity within each task reflects background covariation in ROI activity independent of trial and behavioral factors driving changes in regionspecific activity.
Modularity during each task was calculated from each subject's thresholded (r >= .25), weighted connection matrix using the Louvain method of community detection. This algorithm calculates a global modularity value (Q), reflecting the degree to which a set of ROIs are functioning as distinct modules. PM and AT systems appeared to be functioning as relatively distinct networks during encoding ( Figure 3C), but modularity across our ROIs was significantly reduced during episodic retrieval (t(27) = 3.30, p = .003), suggesting an increase in internetwork communication and a less segregated network structure ( Figure 3B).  Hippocampus was divided into anterior (aHIPP) and posterior (pHIPP). Visualization generated with BrainNet Viewer (Xia, Wang, & He, 2013) . B. Mean change in functional connectivity between encoding and retrieval ('remember') events, including overall modularity as well as between and withinnetwork density (mean strength of connections, defined as r > .25). Bars = Mean +/ 95% CI, points = individual subject mean estimates. * = p < .05. C. Mean ROItoROI connectivity during encoding, retrieval, and retrieval encoding. Connections shown within a task exceed r = .25, p < .05 FDRcorrected, and connections that change between tasks are significantly different from zero, p < .05 FDRcorrected.

Dynamic changes in hippocampalcortical network connectivity predict memory quality
The background connectivity results suggest that episodic retrieval is associated with a less modular hippocampus, PM, and AT network structure, consistent with prior research (Westphal ) . Yet it is unclear whether these changes in network connectivity reflect a general retrieval state or whether they actually support the recovery of complex episodic information. To address this question, we used generalized psychophysiological interaction (gPPI) analyses to measure how effective connectivity of each ROI pair might be modulated by an eventspecific, continuous measure of multidimensional memory quality. This measure captures finegrained information bound in memory, accounting for both the amount and precision of remembered features (see Methods), thus providing a measure of retrieval sensitive to the quality and diversity of memory content. Note that gPPI measures the influence of a seed on a target region after partialling out taskunrelated connectivity and taskrelated activity, and thus the results include an asymmetrical effective connectivity matrix.
Averaging across all possible ROI pairs, as predicted, there was an overall increase in connectivity with eventspecific increases in memory quality (mean beta = 0.36, SE = 0.16; t (27) = 2.17, p = .019). Taking the average of withinnetwork and betweennetwork connections for each seedtotarget pair, we next tested how connectivity across our networks changed with increasing quality of remembered details ( Figure 4A). In line with the results of the background connectivity analyses, it was primarily connections between our networks, particularly with the hippocampus, that increased with memory quality. Specifically, ATPM connectivity increased with higher memory quality (ts ( Figure 4B. Of note, when comparing objects that had been associated with an emotionally negative or neutral sound, increases in network connectivity with memory quality appeared to be slightly stronger for negativeassociated objects, most predominantly for withinPM connections (t(27) = 2.24, p = .033 uncorrected), and ATtoPM connections (t(27) = 2.74, p = .011 uncorrected), although these emotion effects did not survive correction for multiple network comparisons. Exploring these memoryrelated changes in hippocampalcortical network connectivity in more detail, we compared anterior and posterior hippocampus: Is there differential connectivity change with the PM and AT systems along the hippocampal long axis? Comparing the mean of bidirectional connections between each hippocampal subregion and cortical network revealed no differences between aHipp and pHipp, as well as no differences in connectivity change with the AT and PM systems, and no interaction between these factors (Fs(1,27) < 1.96, p s > .17). At the individual region level, there were also no significant differences between pHipp and aHipp in terms of change in connectivity strength with increasing memory quality (|ts| < 2.34, p s > .26, FDRcorrected). Therefore, we found no evidence for differences along the hippocampal long axis; pHipp and aHipp increased their connectivity equally with the cortical systems with higher memory complexity ( Figure 4C).
Finally, we ran two control analyses to test the role of our ROIs in supporting episodic memory quality. First, to determine whether increases in hippocampal synchrony were specific to our networks of interest or whether evident globally, we analyzed wholebrain connectivity changes with memory. Here, we evaluated the main effect of pHipp and aHipp seeds in terms of the modulatory effect of memory quality on seedtovoxel connectivity (see Figure 3D). The hippocampus increased its communication with voxels in a select group of brain regions, . Second, to verify that our ROIs, particularly hippocampus, showed the expected sensitivity to memory retrieval in our task we ran univariate general linear models predicting activity with trialspecific values of memory quality. As expected, mean activity of a number of ROIs, particularly within the PM network and hippocampus, linearly tracked the quality of episodic retrieval ( Figure 4E). Of note, the present connectivity analyses control for changes in regionspecific activity with memory, thus highlighting the additional importance of functional communication of the PM and AT systems and hippocampus to episodic retrieval.

Dissociable PMAT connections predict the precision of recalled item and spatial features
The analyses of multidimensional memory quality provide evidence that changes in PM and AT  composite of the quality of all memory features, it remains unknown how PMAT connections support the fidelity of different types of remembered information. This is particularly important to address in light of existing frameworks that emphasize the role of informational content in determining memory organization (Davachi, 2006; Diana et al., 2007; Eichenbaum, Sauvage, Fortin, Komorowski, & Lipton, 2012; Graham, Barense, & Lee, 2010 . In the medial temporal lobes and connected areas (Ranganath & Ritchey, 2012; Ritchey, Libby, et al., 2015 , AT regions are sensitive to itemspecific associations, and PM regions are sensitive to contextual information, but it is unclear how this organization emerges in terms of network interactions. To this end, we further focused on remember events, specifically trials where a feature was 'successfully' recalled, and tested where changes in connectivity tracked increasing precision of eventspecific i) item color and ii) spatial context, given that these measures were found to be independent in memory (see Behavioral Results).
Looking at the average change in connectivity across every seedtotarget pair, we found that there was an overall positive change in connectivity with the precision of both itemcolor (mean beta = 0.24, SE = 0.11; t(27) = 2.11, p = .022) and spatial memory (mean beta = 0.25, SE = 0.13; t(27) = 1.94, p = .032). Are these overall increases in connectivity driven by distinct patterns? Interestingly, withinsubject correlations between color and spatial ROIxROI gPPI matrices revealed no evidence for a similar pattern in connectivity changes with memory precision for these features (mean z = 0.02, SE = 0.04; t(27) = 0.50, p = .312). At the network level ( Figure 5A), higher color precision in memory was associated with increased connectivity from the AT system to the hippocampus (t(27) = 2.24, p = .017), between the hippocampus and PM system (ts (27)  valence of the object's emotion association (t = 0.32, p = .75). Therefore, itemcolor precision and spatial precision in memory were associated with dissociable network connectivity patterns, and these patterns included an increase in internetwork AT connectivity and hippocampal communication for itemcolor, but an increase in withinPM connectivity and no change in hippocampal communication for spatial information.  Finally, to identify ROI connections that might be contributing to these featurerelated network patterns, analyses were further restricted to focus on 4 seed regions that we hypothesized should show the most representational specificity within our experimental paradigm, including 2 AT regions PRC and AMYG and 2 PM regions PHC and RSC ( Figure 5B). We tested how these seed regions changed their connectivity to all other regions with i) increasing color memory precision and ii) increasing spatial memory precision. All statistics were FDRcorrected.
For color precision, PRC showed the most widespread changes in connectivity to aHipp, PREC, and ANG (ts (27)  connections to these common targets differed by feature ( Figure 5C). The change in connectivity of PM seeds to ANG/PREC was significantly greater for spatial than color precision   (Ranganath & Ritchey, 2012; Ritchey, Libby, et al., 2015 . Here, a PM system provides the spatial contextual scaffold for event details, including item, emotional, and semantic information provided by an AT system. This content is thought to be integrated as an event via the hippocampus (Barry & Maguire, 2018; Moscovitch et al., 2016 . Although a core prediction of the PMAT framework is that functional interactions between cortical systems and the hippocampus are crucial for reinstating multidimensional episodic information, this prediction has not before been tested. First, we found that the PMAT cortical systems functioned in a modular way during memory encoding, with the hippocampus connecting to both systems. In contrast, episodic retrieval was accompanied by a disproportionate increase in internetwork connections. Second, we found that both cortical systems dynamically increased their connectivity to hippocampus with increasing multidimensional quality of episodic memory.
Finally, we found that color and spatial memory precision did not clearly map on to changes in AT and PM connectivity, respectively, but rather that featurerelated differences emerged in how  complements this previous work and serves as a necessary foundation for understanding how the retrieval process alters network dynamics.
Extending evidence of PMAThippocampal integration during retrieval, we found that connectivity between these networks further tracked the eventspecific quality of memory.
Previous research has found that increased functional communication, particularly with hippocampus, seems to be important for 'successful' recollection (Geib, Stanley, Wing, et al., 2017; King et al., 2015; Schedlbauer et al., 2014 . Results of our wholebrain connectivity analysis showed that the hippocampus increased its interaction with a select group of regions, most notably posterior medial and left lateral frontal regions, showing a similar pattern to the results of King et al. (2015) . In prior studies, memory on each trial has been typically quantified in terms of retrieving or forgetting a single episodic feature or a subjective judgment of recollection, thus neglecting the multidimensional quality of event representations. Here, we related network connectivity to a composite score including information about the number of features present in memory as well as quality of those event details. Increased connectivity of PMAT regions to the hippocampus with multidimensional memory quality strongly suggests that hippocampalcortical connections may specifically act to bind multiple sources of information together in memory (Diana et al., 2007; Horner et al., 2015; Ranganath, 2010 , supporting flexible content retrieval (Horner & Doeller, 2017) , rather than simply facilitating access to individual associations or providing a general index of recollection vividness. Therefore, we were able to show that changes in internetwork connectivity parametrically capture the level of detail present in a complex memory representation rather than just the process of retrieval.
Our behavioral results additionally revealed new evidence that episodic memories are bound at the level of the gist of recovered information, rather than the precision with which it is remembered. Retrieving the gist of episodic features showed the expected dependent structure of a hippocampal binding process (Horner & Burgess, 2013 , such that retrieving one feature facilitated memory for the others. This was particularly the case for spatial associations, such that successful retrieval of spatial location was associated with better memory for the other features, supporting the organizational role of space in memory (Robin, 2018) . Interestingly, the precision of each individual feature was at least partially independent of this binding mechanism, such that retrieving a scene location did not significantly improve the precision of color memory, and vice versa, and the precision of recollected item color and spatial context was also unrelated. These results align with the perspective that the primary role of the hippocampus is to bind event features into a coherent spatiotemporal representation but the quality of individual event features occurs at the level of cortical representations (Barry & Maguire, 2018) . Therefore, the precision of bound features is theoretically separable from the binding process itself (cf. The present study design additionally allowed us to examine functional connectivity changes associated with the precision of distinct features within memory. We expected that withinPM and PMhippocampal communication would increase with spatial precision, whereas withinAT and AThippocampal communication would increase with item color precision. Our results partially supported these predictions: internetwork connectivity of the AT system, hippocampus, and PM system tracked the precision of item color memory, whereas connectivity to and within the PM system tracked the precision of scene location memory. In line with our hypotheses, connectivity among PHC, RSC and dorsal PM regions scaled with the precision of spatial but not color memory, suggesting that withinPM connectivity might selectively support the resolution of spatial context associations. Much research has documented the complementary roles role of PHC and RSC in spatial processing and navigation (Epstein, 2008; Mitchell, Czajkowski, Zhang, Jeffery, & Nelson, 2018 , including sensitivity to distance within virtual environments (Sulpizio, Committeri, & Galati, 2014) . Moreover, a recent study showed that RSC is important for forming coherent scene representations similarly using 360° panorama scenes (Robertson, Hermann, Mynick, Kravitz, & Kanwisher, 2016) , in line with its role in viewpoint precision demonstrated here. Surprisingly, we found no evidence that hippocampal connectivity supported the precision of PM spatial representations, which is in contrast to evidence implicating the hippocampus, particularly posterior, in spatial precision specifically (Nadel, Hoscheidt, & Ryan, 2013; Nilakantan et al., 2017, 2018; Stevenson et al., 2018 . In contrast, color precision was associated with connections between AT regions, particularly PRC, to PM regions and hippocampus. Involvement of the PRC complements previous findings that activity of this region is sensitive to item and itemcolor bindings in memory (Diana et al., 2010; Staresina & Davachi, 2008 . However, the finding that itemcolor precision was related to internetwork connectivity, rather than withinAT connectivity, was an unexpected result. There are two possible explanations: First, during episodic reconstruction, the fidelity of item representations may be necessarily integrated within a broader PM contextual framework via the hippocampus. This could explain why hippocampal connectivity supported the precision of color but not necessarily spatial associations in memory. However, color memory precision was not significantly dependent on retrieval of the scene location in our study, providing a tentative argument against this interpretation. Alternatively, the angular gyrus and precuneus may play a contentgeneral role in the retrieval and representation of highfidelity information, thus explaining increased ATPM connectivity associated with itemcolor precision. Previous research has demonstrated consistent involvement of these regions in the representation of subjectively vivid and objectively precise information during memory retrieval using both univariate activation and multivariate methods (Lee et al., 2018; Oedekoven et al., 2017; Richter et al., 2016; Sreekumar et al., 2018 . Moreover, anteriorposterior neural contributions to memory have been proposed to follow a specificity gradient, from gist to precise representations respectively, and not strictly based on informational content (Robin & Moscovitch, 2017) . Our findings lend some support to both perspectives: We find evidence for anteriorposterior content sensitivity in terms of the most influential seed regions, but also common functional projections to angular gyrus and precuneus supporting precise memory retrieval.
The present study revealed network connectivity changes associated with the fidelity of different features during the same retrieval event, indicating that parallel changes in network dynamics support the complexity of episodic memory. Future research should examine the specificity of these corticohippocampal connections more closely, for instance, using causal methods that can adjudicate their specific contributions (cf. Kim et al., 2018; Nilakantan et al., 2017 . These methods will be particularly useful given that episodic memories by definition reflect an integrated structure of item and context information. As such, our data show involvement of ATPM connections, including PRC and PHC seeds, in the precision of both item color and scene location memory, and some prior research has found engagement of PRC and PHC during recollection of both object and spatial information (Burke et al., 2018; Ross, Sadil, Wilson, & Cowell, 2017 . Future research should also account for the temporal evolution of episodic memory, both in terms if the event itself and the retrieval process. For instance,

EXPERIMENTAL MODEL AND SUBJECT DETAILS
28 participants took part in the current experiment (16 females, 12 males). All participants were 1835 years of age (mean = 21.82 years, SD = 3.57) and did not have a history of any psychiatric or neurological disorders. Six additional subjects took part but were excluded from data analyses: two participants did not complete the experiment, one due to anxiety and the other due to excessive movement in the MRI scanner, and four additional participants had chancelevel performance on the memory task (based on criteria outlined in Behavioral Analyses). Informed consent was obtained from all participants prior to the experiment and participants were reimbursed for their time. Procedures were approved by the Boston College Institutional Review Board .
All of the objects were selected on the basis that they did not have a stereotypical color and were also easily recognizable. 120 unique colors from a continuous color spectrum in CIEL*A*B color space were used to change the appearance of the objects, where each color was separated by 3 degrees around a 360 degree spectrum. Each object was resized to 240 x 240 pixels when overlaid on a scene and 300 x 300 pixels when presented alone in grayscale. Six of the IADS sounds accompanying the objects were emotionally negative, as defined by valence rating of less than 4 and an arousal rating of greater than 6 on scales of 1 (low) to 9 (high) from the Bradley and Lang (2007) norms, and had a mean valence of 2.43 (SD = 0.38) and a mean arousal of 7.63 (SD = 0.35). The six neutral sounds were selected to have a valence between 4.5 and 6.5 and arousal less than 5, with a mean valence of 5.31 (SD = 0.42) and a mean arousal of 4.03 (SD = 0.65). All sounds contained natural, easily recognizable content and were 6 seconds in duration. Out of the six panorama scenes used for the experiment, half were indoor locations, including a living room, and office, and a greenhouse, and half were outdoor locations, including a city plaza, a field, and a beach. Each scene was selected through piloting to have no clear areas of symmetry, so that perspectives farther apart, in terms of degrees around panorama, were not obviously more perceptually similar the regions closer together. The original warped panorama images were unwarped to provide naturalistic 100° fieldofview images using the 'pano2photo' function from the SUN 360 database, with each perspective resized to 800 x 600 pixels. Each of the panorama scenes was divided into 120 unique image perspectives, with the center of each perspective shifted by 3 degrees from the previous.

Behavioral Procedure
The experiment was divided into 6 studytest blocks, with all phases completed in the MRI scanner. In each study phase, participants completed 24 trials (see Figure 1A), each of which began with a 1 second fixation, followed by the presentation of an objectscenesound event for 6 seconds. Participants were instructed to remember each object's specific color and location within the panorama scene and were also asked to use the sound to remember the object as a 'bomb' (negative sounds) or as 'safe' (neutral sounds). This instruction encouraged participants to integrate the object and its associated features into a meaningful event. Within a study block, each panorama scene was shown 4 times and each sound was encoded twice. All objects were trial unique. The object color and scene location values were pseudorandomly selected with the constraint that objects associated with the same panorama within the same block should be at least 45 degrees apart in their color and location within the scene to minimize interference. The trial order was randomized within each block for every participant. Therefore, across the experiment, participants studied 144 objectscenesound events, with 72 objects accompanied by negative sound and 72 accompanied by a neutral sound, and 24 objects associated with each of the six panorama scenes. Allocations of the objectcolorscenesound associations were randomly generated for each subject.
In each test phase, participants were tested on their memory for all 24 encoded events. On each trial, a grayscale version of a studied object was shown for 4 seconds. During this time, participants were asked to recall all of the details associated with that object during the study phase (emotion association, color, and scene location) and to hold that whole image in mind as vividly as possible (see Figure 1B). Participants then had an additional 2 seconds to indicate the object's emotional association. Following a 1 second fixation, participants were then shown the objectscene pairing that they studied, but the object was presented in a random color, in a random location of the associated panorama scene. Participants were asked to reconstruct both the color and scene location of the object as precisely as they could, the order of which was counterbalanced across trials. Participants had up to 6 seconds to reconstruct each feature, with a 1 second fixation separating these questions. For the 'color' question, participants were instructed to use two button box keys to move counterclockwise or clockwise around the color spectrum to find the color of the object as they studied it originally (target color). For the scene question, participants were asked to move counterclockwise or clockwise around the panorama to find the location in which the object was originally presented (target scene location). The feature value that participants chose for the first question was carried over to the second question. At the end of each test phase, participants were presented with feedback on their performance for 12 seconds, including the percentage of the time they correctly identified objects as bombs or safe, and the percentage of the time that they were 'close' (defined as +/ 45 degrees from the target feature value) to the original color or scene location of the objects.

Functional Neuroimaging Data Acquisition
MRI scanning was performed using a 3T Siemens Prisma MRI scanner at the Harvard Center for Brain Science, with a 32channel head coil. Structural MRI images were obtained using a T1weighted (T1w) multiecho MPRAGE protocol (field of view = 256 mm, 1 mm isotropic voxels, 176 sagittal slices with interleaved acquisition, TR = 2530 ms, TE = 1.69/ 3.55/ 5.41/ 7.27 ms, flip angle = 7º, phase encoding: anteriorposterior, parallel imaging = GRAPPA, acceleration factor = 2). Functional images were acquired using a whole brain multiband echoplanar (EPI) sequence (field of view = 208 mm, 2 mm isotropic voxels, 69 slices at T > C25.0 with interleaved acquisition, TR = 1500 ms, TE = 28 ms, flip angle = 75º, phase encoding: anteriorposterior, parallel imaging = GRAPPA, acceleration factor = 2), for a total of 466 TRs per scan run. Fieldmap scans were acquired to correct the EPI images for signal distortion (TR = 314 ms, TE = 4.45/ 6.91 ms, flip angle = 55º). Physiological data, including heart rate and respiration, were also collected but were not further analyzed.  (Esteban et al., 2017) was used as a preliminary check of MRI data quality. Scan runs were excluded from data analyses if more than 20% of TRs exceeded a framewise displacement of 0.3 mm. Two participants had 1 scan run excluded using this threshold. A further four participants also successfully completed only 5 out of the 6 scan runs, 3 as a result of exiting the scanner early and 1 due to a technical problem with the sound system during the first run.
All data preprocessing was performed using FMRIPrep v1.0.3 (Esteban et al., 2018) with the default processing steps. To summarize: each T1w volume was corrected for intensity nonuniformity and skullstripped. Brain surfaces were reconstructed using reconall from

Behavioral Analyses
Participants' responses for the item color and scene location questions were analyzed by fitting a mixture model (Bays et al., 2009; Zhang & Luck, 2008

Functional Connectivity Analyses
All connectivity analyses were conducted using the CONN toolbox (WhitfieldGabrieli & NietoCastanon, 2012) . In all cases, functional data were first denoised within each scan run, including demeaning, linear detrending, highpass filtering at 1/128 Hz, and regression of the first principal component from aCompCor to remove white matter and CSF confounds , framewise displacement, and 6 motion parameters. All connectivity estimates were then calculated across the concatenated functional runs, as is standard in CONN. All analyses for hippocampus, PM and AT ROIs used unsmoothed functional data to ensure no voxels were included in mean estimates from outside these anatomical regions. Whole brain analyses used functional data smoothed with a 5mm FWHM gaussian kernel, masked by gray matter.
Connectivity estimates were calculated between the mean time series of each bilateral ROI and then averaged at the network level where applicable.

Task Background Connectivity
For analyses of network dynamics during encoding and retrieval, we ran a background connectivity analysis. Here, we first created two task vectors reflecting the occurrence of i) encoding and ii) 'remember' events during the functional time series. Each event was modeled as a HRFconvolved delta function and all other time points were assigned a value of zero. Five additional parametric covariates were generated for each event type to capture memory effects during encoding and retrieval: emotion memory, where trials were coded as incorrect (0), low confidence correct (0.5), or high confidence correct (1), color and scene retrieval success, reflecting binary correct (1) vs. incorrect (0) retrieval, and the precision of 'correct' color and scene memory, coded as the reversescored error of remembered trials. Regressors for emotion memory and color and scene retrieval success were meancentered across all trials within an event (encoding or 'remember'). Regressors for color and scene precision were meancentered within all successfully remembered trials for that feature. As with the task regressors, all other time points within these memory covariates were then set to zero and the vectors were convolved with the HRF. All task effects and memory covariates were regressed out from the functional data prior to connectivity analyses as part of CONN's denoising step. Therefore, results represent connectivity during encoding and retrieval tasks independent of trial and memoryrelated changes in region activity.
To measure connectivity between our ROIs during encoding and retrieval, we calculated the Pearson's correlation between each pair of mean ROI time series weighted by the vectors indicating encoding and remember events. This produced two 12x12 correlation matrices for each subject one per task. We computed 3 measures to compare background connectivity between episodic encoding and retrieval within each subject: 1) Modularity, reflecting the degree to which our regions were operating as distinct networks, computed using the Louvain algorithm from R's NetworkToolbox (Christensen, 2018) . This method calculates a global modularity value

MemoryModulated Connectivity
Generalized psychophysiological interactions (gPPI) analyses (McLaren, Ries, Xu, & Johnson, 2012) were used to investigate changes in network connectivity with memory performance from trialtotrial during remember events. Two models were constructed. The first model tested the modulatory effect of an objective measure of 'multidimensional memory quality' on connectivity.
To create this composite measure, memory for each feature (emotional association, item color, scene location) was scaled between 0 (incorrect) and 1 (perfect recollection) on each trial. Low confidence, correct memory for the emotion was coded as 0.5 and correct, high confidence emotion memory was coded as 1. Correct memory for the color and scene features was scaled according to precision, where a value of 1 would reflect perfect feature memory (an error of 0).
These values were summed so that each trial could have a total memory quality score between 0 and 3, with higher values reflecting better memory. Therefore, a maximum value is achieved on any given trial not by simply remembering all features, but by remembering them all with perfect precision. This memory quality vector was meancentered within remember events and convolved with the HRF, with all other time points set to zero. For gPPI analyses, the mean time series of each ROI was predicted by the mean time series of a seed region, a psychological variable containing the HRFconvolved memory quality scores, and the interaction of the seed time series and memory regressor. Taking these interaction terms produced a 12x12 gPPI matrix for each subject reflecting the change in functional connectivity from each seed to target region with higher memory quality (e.g., stronger seedtarget connectivity when memory quality is high compared to low). Note that as gPPI measures the taskrelated change in influence of a seed on a target region after partially out taskunrelated connectivity and taskrelated activity, the outcome is an asymmetrical effective connectivity matrix.
We then tested whether changing network connectivity might be related to the precision of specific features in memory. In a second model, 5 parametric modulators captured memory retrieval and precision for the individual episodic features during 'remember' events, as described in Task Background Connectivity: emotion memory (coded in terms of incorrect, low confidence correct, high confidence correct), color and scene retrieval success (coded as binary correct vs. incorrect retrieval), and the precision of 'correct' color and scene memory, coded as the reversescored error of remembered trials. In this gPPI analysis, each target ROI time series was predicted by a seed time series, all 5 memory regressors, and the 5 seed*memory