Memory hierarchies map onto the hippocampal long axis in humans.

Memories, similar to the internal representation of space, can be recalled at different resolutions ranging from detailed events to more comprehensive, multi-event narratives. Single-cell recordings in rodents have suggested that different spatial scales are represented as a gradient along the hippocampal axis. We found that a similar organization holds for human episodic memory: memory representations systematically vary in scale along the hippocampal long axis, which may enable the formation of mnemonic hierarchies.

The hippocampus, a region critical for memory as well as internal representation of space[1], differs in structure and function along its long-axis (dorsal-ventral in rodents, and posterior-anterior in humans)[2-4]. More specifically, encoded space increases in scale along the long-axis of the hippocampus, as evidenced by an increase in place-field size in the rodent hippocampus from dorsal to ventral hippocampus[5]. Such a gradient may provide a mechanism, which enables multiple scales of episodic memories, ranging from detailed individual events to more comprehensive multi-event narratives, to be concurrently represented along the long-axis of the hippocampus, and may underpin hierarchical memory representations[6,7]. However, whether a similar neural gradient indeed underlies organisation of episodic memories in humans remains unclear[5,8,9].

To test if memory scale is differently represented along the hippocampal long-axis, we examined the emergence of novel, multi-event representations by combining fMRI and multivoxel pattern analysis with a ‘Narrative-insight task’ (Fig. 1, and Supplementary Fig. 1)[10]. We used realistic, life-like videos showing individual events, which could be integrated into narratives[10] and thus experimentally simulate processes involved in episodic memory formation. We gradually built up four separate narratives by initially presenting seemingly unrelated events (A, B, C and X in Fig. 1), before sequentially introducing two different ‘linking’ events (first L1, then L2), which provided insight into direct (A-B, and B-C, respectively) and inferred (A-C) event-associations within the narratives. Memories of such narratives can be recalled at different resolutions ranging from detailed events to more comprehensive narratives. Thus, although all these events were part of the same narrative, we hypothesized that the representation of such a multi-event narrative may differ along the long-axis of the hippocampus depending on the ‘narrative scale’ of the ensuing representation. We propose that there are different resolutions in which these narratives can be represented, which we refer to as small-, medium-, and large-scale networks (see Fig. 2a). Here, a representation at the largest narrative scale would reflect the complete narrative (large-scale network), including all possible event associations between both directly (A-B and B-C) and indirectly (A-C) related events in Phase 3 of the task, and would be represented in the anterior hippocampus. At the smallest narrative scale, the representations would contain only individual event-pair associations, which would, in this context, reflect only the most recent directly associated event pairs (small-scale network) and would be represented in the posterior hippocampus. An intermediately scaled representation may concurrently contain multiple event-pair associations in Phase 3 of the task, but would not bridge between them (medium-scale network), and would be represented in the mid-portion of the hippocampus.

Figure 1

Schematic overview of the narrative-insight task

Participants were presented with animated videos of life-like events. Videos of one narrative were presented in 5 phases. Phase 1, 2, 3 contained events A, B, C and control-event X. Link-phases 1 and 2 were interleaved during which events L1 and L2 were presented and provided links between events A and B, and events B and C, respectively. Thus, some events were directly linked (first A and B via L1, then B and C via L2), while other associations had to be inferred (A and C were associated via their shared association with B). Participants performed 4 runs; in each run a different narrative was presented (narrative 1, 2, 3 and 4).

Figure 2

Increasing memory scale along the hippocampal long-axis

(a) Depiction of the three network scales. Contrast weights for the small-scale network, separately for phase 1, 2 and 3: A-B: -1 2 -1 / B-C: -1 -1 2 / B-X: 2 -1 -1; cf. schematic bars); the medium-scale network for phase 2 and 3: A-B: -1 1 / B-C: -1 1 / A-C: 1 -1 / B-X: 1 -1; and the large-scale network for phase 2 and 3: A-B: -1 1 / B-C: -1 1 / A-C: -1 1 / B-X: 3 -3. Note that these three networks do not correspond to experimental phases. We predict a mnemonic-scaling contrast which reflects an interaction between Narrative scale and hippocampal ROI: (Small-scale, Medium-scale, Large-scale: pos: 2 -1 -1, mid: -1 2 -1, ant: -1 -1 2). (b) Model evidence (parameter estimates) of left and right hippocampus (mean ± S.E.M.) separately for the three ROIs and scales (N=29). Small-scale network: a representation of the narrative sensitive only to the directly linked events immediately after linking (link between event A and B replaced by re-linking B to C later in time) was observed in the posterior hippocampus only (posterior: F1,28 = 4.1, approaching significance at p = 0.053; mid-portion: F1,28 = 0.002, p = 0.96; anterior: F1,28 = 0.05, p = 0.82). Medium-scale network: Increase in neural similarity between both pairs of directly linked events simultaneously, relative to inferred link and control-event X, was present only in the mid-hippocampus (mid-portion: F1,28 = 4.99, p < 0.05; anterior: F1,28 = 0.34, p = 0.56; posterior: F1,28 = 3.14, p = 0.09). Large-scale network: the anterior hippocampus showed a selective increase in neural similarity between all three events (A-B, B-C, A-C) within each narrative, in contrast to X (anterior: F1,28 = 8.6, p < 0.01; posterior: F1,28 = 1.96, p = 0.17; mid-portion: (F1,28 = 1.63, p = 0.21). (c) Depiction of the three ROIs.

We employed representational similarity analysis (RSA), which uses correlations of across-voxel activation patterns as a proxy of neural similarity, to quantify memory representations along the long-axis of the hippocampus. We split the hippocampus into three regions of interest (ROIs), an anterior portion, a mid-portion and a posterior portion (see Online Methods) and computed correlation coefficients between across-voxel activation patterns within each ROI for event-pairs of interest (A-B, B-C and A-C) in each of the three experimental phases separately and averaged effects across the 4 runs (see Online Methods, and Supplementary Fig. 1). B-X served as a control (see Supplementary Fig. 2 for more information). We tested the predicted interaction effect using a reduced a priori model, referred to as ‘mnemonic-scaling contrast’. Using this contrast (with contrast weights for Small-scale, Medium-scale and Large-scale: posterior: 2 -1 -1, mid-portion: -1 2 -1, anterior: -1 -1 2; see Online Methods and Fig. 2a) in a repeated-measure ANOVA, with Narrative scale (small, medium, large), ROI (posterior, mid-portion, anterior), and Hemisphere (left, right) as within-subject factors, we found a significant interaction effect between Narrative scale and ROI (F1,28 = 11, p < 0.01; see Fig. 2). Thus, the small-scale network was indeed represented in the posterior portion, the medium-scale network in the mid-portion and the large-scale network in the anterior portion of the hippocampus. There was no difference between hemispheres (Supplementary Fig. 3). Additional control analyses showed that these results were unlikely to be due to an increasing amount of information or passing of time throughout the task (see Supplementary Fig. 2), or BOLD signal fluctuations (see Supplementary Fig. 4).

Post-hoc analyses revealed evidence for the smallest scale of narrative representation only in the posterior portion of the hippocampus (Fig. 2 and Supplementary Fig. 3), which is consistent with our previous report[10]. In addition, the medium-scale network was only represented in the mid-portion of the hippocampus, suggesting co-existence of the two directly integrated event-pair associations, without bridging across the inferred associations[11,12]. Finally, we found that only the anterior portion of the hippocampus showed evidence for large-scale networks, which included the two directly integrated event-pair associations and a bridge across the inferred association not directly experienced. The crucial difference between the large-scale network and the medium-scale network is that the former predicts this inferred association between A and C, an effect that was restricted to the anterior hippocampus (Supplementary Fig. 5).

These data suggest that multi-event narratives are simultaneously represented at multiple narrative scales along the hippocampal long-axis, but is this gradient relevant for behavior? Following scanning, participants were asked to report all narratives they have seen during the experiment. The results indicated that some participants remembered 4 unified narratives (12 participants), while others failed to integrate some events into unified narratives (Supplementary Fig. 6). We used this difference in performance to split the participants into a ‘full-integration’ and a ‘partial-integration’ group, and investigated whether the gradient is expressed differently between those groups (see Online Methods). We observe that the long-axis gradient was only present in the full-integration group (group-by-scale-by-ROI interaction: F4,104 = 2.7, p < 0.05; Supplementary Fig. 6 and 7). These results suggest that representing event associations at multiple scales simultaneously supports memory recall of accurate integrated narratives.

In sum, here we provide the first evidence in humans that event associations are represented as memory hierarchies with multiple associative networks increasing in scale along the hippocampal long-axis: small-, medium-, and large-scale networks are represented in posterior, mid-portion, and anterior hippocampus, respectively. Moreover, this hierarchical memory gradient is related to accurate recall or construction of integrated narratives. These results demonstrate that a mnemonic gradient underlies organisation of human episodic memory, which may relate to the gradient of the scale of encoded space[5].

Previous research showed involvement of the mid-portion and anterior hippocampus during inference that could be driven by bridging between unseen associations or the formation of more complex networks, potentially with more complex networks represented anteriorly and less complex networks represented in mid-portion[11,13] (Supplementary Fig. 10). One possibility is that the large-scale network effect in the anterior hippocampus reflects such a bridging function. Alternatively, it might index a complete and integrated representation akin to a relational memory network[1,7] or cognitive map[14]. Importantly, these two explanations are not mutually exclusive; making inferences about unseen connections is crucial for the creation of large-scale mnemonic networks. This dovetails with previous research that showed a functional dissociation within the rodent hippocampus: ventral hippocampal neurons in rats represent global event context[15] while neurons in dorsal hippocampus encode more specific event information[15,16].

Our results accord with previous findings on the role of the hippocampus in memory generalisation. For example, novel conceptual knowledge[17] and the formation of schemas[18] require certain degree of abstraction from individual events and seem to preferentially involve anterior hippocampus. In contrast, smaller networks consisting of few elements seem to engage more posterior regions[10]. There are many different proposals about hippocampal long-axis functional dissociation (see Supplementary Fig. 8). The hippocampal memory gradient may provide a mechanism which enables multiple scales of episodic memories, ranging from individual events to more comprehensive multi-event narratives to be represented by the brain simultaneously as different levels of mnemonic hierarchies[6,7]. An interesting question for future research relates to whether each level of the memory gradient is specific to one scale of the narrative representation, or alternatively, if anterior hippocampal sub-regions extend the more posterior lower scale narrative representations.

Although it is clear from the present data that this scaled coding mechanism supports memory performance, the question of how the hierarchical representation relates to performance remains unanswered. One possibility is that overall memory benefits from both maintaining the ability to individuate memories of separate events[19] and to integrate multiple experiences for the purpose of generalisation or abstraction of knowledge[17,20]. Representing events at multiple scales may provide an effective way of providing a context or schema for individual events, which is known to improve memory performance and may protect against loss of individual event details.

In conclusion, we provide evidence for a mnemonic gradient along the hippocampal longitudinal axis, which enables the concurrent representation of multiple memories in hierarchies, a finding with potential implications for the classroom.

ONLINE METHODS

Participants

Thirty-five healthy students from the Radboud University campus (17 males) participated in this study. All participants were right-handed. Six participants were excluded from further analyses: four due to excessive head motion (> 2mm); and further two due to technical problems leading to incomplete data sets. The final group consisted of 29 participants (14 males, aged 18-33 years, mean age 22.8) who all had normal or corrected-to-normal vision. All participants gave written informed consent. The study was approved by the local ethics committee (CMO Arnhem-Nijmegen, The Netherlands).

Study design

Narrative-insight task (scanning)

First, participants completed the narrative-insight task[21] in the MRI scanner. Stimuli consisted of four animated narratives generated using The Sims 3 game (www.thesims3.com). Each narrative consisted of five separate events of 5 seconds in duration. Importantly, participants were presented with only one narrative per scanning run, with 4 scanning runs in total. Additionally, each run contained one control event (event X). New information was introduced twice in each run so that the events were gradually integrated into one narrative (Fig. 1).

Each run consisted of five different phases (see example narrative in Fig. 1). During phase 1, 2 and 3, participants saw 4 seemingly unrelated events: A (e.g. a man eating soup), B (e.g. a child playing on the floor), C (e.g. a man watching TV) and X (e.g. a man cooking). Between phase 1 and 2, a new event was presented (event L1), which linked two of the seemingly unrelated events: A and B (e.g. the man from event A brings the child from event B to bed). A similar linking event (event L2) was presented between phases 2 and 3, and linked another two of the seemingly unrelated events: B and C (e.g. the man from event C gives a bottle to the child from event B). Therefore, by the third phase, it was clear to the participants that events A, B and C were all part of the same narrative with direct links between A and B, and B and C, and an indirect link between A and C (via B). The content of L1 and L2 was counterbalanced between subjects. Since the X-event was never linked to any of the other events, it served as a control.

A pseudo-randomized order was used to present each event in phase 1, 2 and 3, i.e. all four events were shown before an event was repeated and an event was never shown twice in a row; this was done to avoid temporal confounds in the representational similarity analysis (RSA). Each event was shown six times per phase, with an inter-trial interval of 5.3 seconds on average (1, 4 or 11 seconds, uniform distribution). Thus, each original event (A, B, C, X) was presented 18 times in the task in total. The link-events were repeated six times as well, interspersed with inter-trial intervals of 12 seconds on average (6, 12 or 18 seconds, uniform distribution), see Supplementary Fig. 1 for a schematic overview of the task structure. Participants finished the entire task structure for one narrative (Phase 1, link-event 1, Phase 2, link-event 2, Phase 3) before continuing with the same task structure for the following narrative.

Additionally, there was a target event (a girl riding a scooter), not related to any of the narratives, which was presented at random moments during the experiment (in 8% of all trials). Participants were instructed to press a button whenever they saw this target event. The purpose of the target event was to make sure that participants were attending the stimuli. Before the first run, participants were presented with an example narrative (shown in Fig. 1). All example events were shown twice following the same procedure as in the actual task narratives to ensure that participants understood the task well.

The task was presented using Presentation software (Neurobehavioral Systems, version 16.2). Afterwards, participants were taken out of the MRI scanner and completed a narrative-recall task, in which they were asked to write down concisely all narratives they have seen during the task (20 minutes). In this within-subject design, no blinding procedures were applied for data collection and analysis.

Additional behavioral experiment

Since behavioral testing was done after the Narrative-insight task was completed, we ran a separate behavioral experiment to test participants’ memory performance immediately after the first link (see Supplementary Fig. 9 for more details).

MRI acquisition

All images were acquired using a 3T TIM Trio scanner equipped with a 32 channel head coil (Siemens, Erlangen, Germany). For the functional images, a 3D Echo Planar Imaging (EPI) sequence was used[22], with the following parameters: 56 axial slices, voxel size 1.5mm isotropic, TR = 1888 ms, TE = 26 ms, flip angle = 16 deg, GRAPPA acceleration factor = 2, FOV = 222*222*84 mm. In addition, a structural T1 sequence (MPRAGE, 1mm isotropic, TE = 3.03 ms, TR = 2300 ms, flip angle = 8 deg, FOV = 256*256*192 mm) was acquired. A dual echo 2-D gradient echo sequence with voxel size of 3.5 × 3.5 × 2.0 mm, TR = 1020 ms, dual echo (10 ms, 12.46 ms), flip angle = 90 deg, and separate magnitude and phase images was used to create a gradient field map to correct for distortions.

MRI data analysis

Preprocessing

Image preprocessing was performed using the Automatic Analysis Toolbox[23], which uses custom scripts combined with core functions from SPM8 (www.fil.ion.ucl.ac.uk/spm), FreeSurfer (http://surfer.nmr.mgh.harvard.edu/) and FSL (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/). The structural images were bias-corrected and de-noised using an optimized non-local means filter to improve image quality[24]. The unified realign and unwarp procedure, as implemented in SPM8[25], was used to correct for head motion and voxel displacement due to magnetic-field inhomogeneity. Co-registration of the functional images with the structural image was performed with the following procedure: the structural image was co-registered to the T1 template, and the mean EPI was co-registered to the EPI template. This co-registered mean EPI was then co-registered to the structural image. The co-registration parameters of the mean EPI were applied to all functional volumes. Functional images were corrected for physiological noise with RETROICOR[26]. RETROICOR uses 10 cardiac phase regressors (5th order fourier set), 10 respiratory phase regressors (5th order fourier set), and 6 other nuisance regressors (heart rate frequency, heart rate variability, raw respiration data, respiratory amplitude, respiratory frequency, and frequency times amplitude of respiration). The FSL brain extraction toolbox was used to create a skull-stripped structural image. This structural image was segmented into grey matter (GM), white matter (WM) and cerebro-spinal fluid (CSF) with SPM8[27]. Mean-intensity values at each time point were extracted for WM and CSF and used as nuisance regressors in the GLM analysis (see below).

Physiological measures

To correct for physiological noise (see above for details), heart rate was monitored with a pulse oximeter placed on the ring finger of the left hand using BrainAmp (ExG amplifier, Brain products GmbH). Participants were instructed to keep this hand as still as possible during the experiment. Heart rate data were inspected and corrected for movement-related and other measurement artefacts. Respiration was recorded at a sampling rate of 400 Hz using the respiration belt enclosed by BrainAmp (E×G amplifier, Brain products GmbH).

First-level modelling

For each narrative separately, each event in phase 1, 2 and 3 of the experiment (event A, B, C, X) was modelled with a general linear model (GLM) using two separate regressors: one for the 3 odd trials and one for the 3 even trials. These regressors were convolved with the canonical haemodynamic response function (HRF). First-level modelling was performed according to the methods suggested by Mumford et al[28]. In short, for each regressor of interest, a separate GLM was performed containing the regressor of interest and another regressor including all other events of the experiment. This resulted in 96 GLMs per participant, with 2 regressors for each event (A, B, C, X) in each phase (1, 2, 3) for each narrative. Additionally, each GLM included: target events, link-events (link 1 and 2), and button presses (all convolved with the HRF), and 6 motion parameters (translations of X, Y, and Z coordinates, pitch, roll, and yaw), mean signal intensity in CSF, mean signal intensity in WM, and 26 regressors for physiological noise (see preprocessing for more detailed explanation). High pass filtering with a cutoff of 128 seconds was used to remove effects of low-frequency signal drifts.

Representational Similarity Analysis (RSA)

We defined a priori regions of interest (ROIs, see ‘ROI definition’ below) and examined the correlation between across-voxel activation patterns of first-level beta estimates within these ROIs as a proxy of neural similarity[29]. The regressors modelling odd and even trials for events A, B, C and X were considered as the regressors of interest. We averaged across the correlations for odd and even trials which led to a 48 by 48 matrix of correlations (4 events of interest per phase, 3 phases per narrative, and 4 narratives). Only event-pair correlations for event-pairs within the same task phase were analyzed (see Supplementary Fig. 1). These Pearson’s correlation coefficients were normalized using Fisher Z transformation. We then defined contrasts designed to model three different representational networks (see Supplementary Fig. 1 for details):

Small-scale Network

We predicted that the smallest narrative scale would contain only representations of individual event-pair associations, which would, in this context, reflect only the most recent directly associated event pairs and would depend on the posterior portion of the hippocampus. The contrast we used to test this prediction reflected an increase in A-B similarity from phase 1 to phase 2 followed by a decrease again from phase 2 to phase 3, combined with an increase in B-C similarity from phase 2 to phase 3, relative to B-X similarity. The contrast weights for phase 1, phase 2 and phase 3 were as follows: A-B: -1 2 -1 / B-C: -1 -1 2 / B-X: 2 -1 -1; cf. schematic bars in Fig. 2A.

Medium-scale Network

We predicted that a ‘medium-scale network’ would concurrently represent multiple event-pair associations, but without bridging between them. The contrast we used to test this prediction reflected an increase in A-B similarity and B-C similarity between phase 2 and phase 3, relative to A-C similarity and B-X similarity, with the following contrast weights for phase 2 and phase 3: A-B: -1 1 / B-C: -1 1 / A-C: 1 -1 / B-X: 1 -1, see Fig. 2A.

Large-scale network

Finally, the ‘large-scale network’ includes all possible event associations between both directly (A-B and B-C) and indirectly (A-C) related events. The contrast we used to test this prediction reflected an increase in A-B similarity, B-C similarity, and A-C similarity, between phase 2 and phase 3 relative to B-X similarity. To determine whether this model indeed reflected presence of indirect associations, we looked at an increase in similarity of the indirect link (A-C) separately, relative to control B-X similarity, with the following contrast weights for phase 2 and phase 3: A-B: -1 1 / B-C: -1 1 / A-C: -1 1 / B-X: 3 -3; see Fig. 2A.

Contrasts were normalized by dividing each contrast by the square-root of the sum-of-squares of its contrast weights, which permitted us to directly compare the parameter estimates of the three different models. The sample size was based on our previous study[10] and power analysis was performed with G*power (dz=1.033, alpha=0.0001, power=0.8). Participants were not grouped and therefore no randomization of participants was performed.

ROI definition

A hippocampal mask was constructed using the WFU pickatlas[30]. We predict a gradual change along the long-axis of the hippocampus, and therefore split the hippocampal mask in approximately equal lengths along the long-axis (posterior portion of the hippocampus: from Y = −40 to −30; mid-portion of the hippocampus: from Y = −29 to −19; anterior portion of the hippocampus: from Y = −18 to −4). A unified segmentation procedure (SPM8) was used to estimate parameters relating individual anatomy to MNI space. The inverse normalization parameters were used to create subject specific (grey-matter) ROIs in native space based on the MNI masks described above (since first-level modeling was performed in native space).

ROI analyses

A repeated-measures ANOVA with a mnemonic scaling contrast (i.e. a reduced a priori model with predicted contrast for small-scale, medium-scale and large-scale network, respectively: posterior: 2 -1 -1, mid-portion: -1 2 -1, anterior: -1 -1 2) with within-subject factors Scale (small, medium, large), ROI (posterior portion, mid-portion, anterior portion), and Hemisphere (left, right) was used to test the prediction of the increasing gradient (small-scale in posterior portion, medium-scale in mid-portion, large-scale in anterior portion of the hippocampus). To investigate this gradient further, we performed post-hoc repeated-measures ANOVAs for each model (small-scale, medium-scale, large-scale) and each ROI (anterior portion, mid-portion, posterior portion) separately with Hemisphere (left and right) as within-subject factor, see above and Fig. 2a for details. To examine the interaction between the fMRI and behavioral results, we performed a repeated-measures ANOVA with within-subject factors Scale (small-scale, medium-scale, large-scale), ROI (anterior portion, mid-portion, posterior portion), and Hemisphere (left, right), and included ‘Performance group’ as between-subjects factor based on the performance in the Narrative-recall task (full integration of all 4 narratives versus partial integration).