Abstract
For more than a century, experimental research on human memory has focused on individuals learning and remembering in isolation; memory scientists began a study of social influences in earnest only in the last 3 decades. A key phenomenon driving this research is collective memory, or defined in cognitive terms, memories that a group of people share. While attention has focused on the overlap in the contents of memory, we focus on deeper representations of these shared memories, namely overlap in their recall organization. Individual memory research has extensively examined how memories become organized, but this is a new arena of research on social remembering. We describe two novel applications of quantitative tools that capture at a global level the overlap in how people organize the contents of their collective memory. These two tools provide different approaches for capturing the ways in which elements become organized and give us holistic views of how people represent the interconnections in memory for a particular episode or theme. This initial research assessing the development of overlapping memory organization offers a stepping stone toward understanding how collective memory narratives and schemata emerge and potential consequences for future learning.
Keywords
Collective memory has been of long-standing interest in the humanistic and social sciences, in which this concept is typically defined as memories that members of a group share and that are related to their identity (Assmann & Czaplicka, 1995; Halbwachs, 1980). In psychological science, particularly in cognitive experimental research, a study of collective memory is more recent. In these empirical approaches, the role of identity features to different extents in characterizing collective memory, and definitions vary in relation to the aspect of collective representations being captured. For example, these definitions include collective memory as a body of knowledge, as a schema or attribute of past events, or as a process of remembering and forgetting (see Roediger & Abel, 2015).
In psychological experimental research, a reference to social identity has not been the prevalent practice, and these studies usually define collective memory without reference to identity, including experiments that use narratives as study-recall materials (e.g., Coman et al., 2016; Cuc et al., 2006; Weldon & Bellinger, 1997). Within this realm, we have focused on understanding the cognitive psychological underpinnings of how people form collective memory (Rajaram & Pereira-Pasarin, 2010). This approach adds unique value by showing how information sharing per se leads to the development of shared memories and, in turn, informs the construction of collective memory defined more broadly.
In this article, we approach the study of collective memory from this cognitive perspective and focus on two definitions: a shared representation that emerges from a process of collaborative recall of the past by a group and a body of knowledge that is shared by many people. The former notion of collective memory emerges from a process of collaborative remembering in which members of a group work together to recall information they encountered earlier that activates many cognitive processes to shape recall (Basden et al., 1997; Rajaram & Pereira-Pasarin, 2010; Weldon & Bellinger, 1997; see also Bietti & Sutton, 2015). This is typically studied in laboratory-based experiments (Greeley & Rajaram, 2023). The latter notion, collective memory as a body of knowledge, does not rely on information sharing within the research setting; instead, it captures overlapping information accrued by a group as relating to their shared context. This is often measured with survey questions and similar tasks (e.g., Abel et al., 2019; DeSoto & Roediger, 2019). The two studies we discuss in this article follow these respective methodologies for harnessing data on collective memory.
Irrespective of methodology, memory researchers interested in empirical approaches to studying collective memory typically focus on what people remember, that is, the overlap in the contents of memory across people. In this article, we focus on a less studied but fundamentally important aspect of memory overlap—how people organize the information they recall (Healey & Kahana, 2014; Mandler, 1967). Specifically, we consider the extent to which this recall organization is synchronized across members of a group, that is, the nature of collective memory organization.
Why study collective memory organization? In considering this question, a tapestry serves as a useful analogy. Two tapestries may contain elements with the same shapes and colors. However, in one tapestry these shapes and colors may be arranged in a haphazard fashion, whereas in another tapestry they may be arranged to create a scene such as a lake with surrounding trees nestled at the base of a mountain. Now consider a process that moves the elements of the two tapestries to similar positions such that both acquire a similar overall pattern. These two tapestry patterns still have roughly the same set of shapes and colors, that is, similar contents, but they now also have similar arrangements of these elements, that is, similar organizations.
The tapestry example is analogous to the way we may construct memory narratives to describe past episodes, experiences, and conversations. Collaborative recall can facilitate remembrance and forgetting of the same details from a past episode or conversation. While these overlapping details in memory for those who collaborated give rise to collective memory content, similarity in the arrangement of these details within the recall sequence creates synchrony in the way recall becomes organized across collaborators. Two narratives could share many overlapping ideas and details, but they could be packaged quite differently; certain information could be associated differently across narratives, recalled earlier or later relative to its position in the original version, or discarded altogether. These alterations can change how the content becomes organized and the future renditions of these narratives.
Consider here Bartlett’s (1932) classic case studies of repeated and serial reproduction in recall. In repeated reproduction, the same individual recalled a story repeatedly, changing it with each attempt. In serial reproduction, one individual’s story recall became study information to the next individual, who then recalled it, and so on, with dramatic changes occurring over social transmission. A striking example Bartlett reported concerned the recall of the Native American story “War of the Ghosts” that in the hands of his British participants altered over recall attempts to align to more local flavors. Bartlett’s (1932) examples showed how retelling can change not just the details people remember (or forget) but the positional changes of these details within the narratives. Although Bartlett’s examples did not venture into group recall, these examples taken together invite systematic investigations into how social remembering can shift narrative structures.
In addition to helping us understand shifts in memory narratives, what might be other potential consequences of aligned memory structures? A significant body of research on individual recall offers some pointers, particularly with respect to the impact of cues on influencing recall, the retrieval dynamics during recall, and test-potentiated new learning (TPNL). Retrieval cues (or, during collaboration, others’ responses) can help or hurt recall depending on when they occur (Basden et al., 1997). Depending on whether memory structures are different or aligned, cues could change memory performance. With respect to retrieval dynamics (Greeley & Rajaram, 2023; Healey & Kahana, 2014), synchronized structures can lead to similar retrieval strategies for the sequence of recall and for learning new information that fits within the memory structures. With respect to TPNL, the finding that the retrieval practice of “X” benefits the subsequent learning of “Y” compared with restudying “X” is typically attributed to context change. But some evidence suggests that retrieval practice provides an opportunity to update strategies in a way that is beneficial for subsequent learning (Chan et al., 2018). Thus, it is possible that former collaborators will experience similar downstream learning outcomes.
Our research takes the initial key steps toward ultimately addressing whether aligned memory structures can produce these potential memory consequences. We begin with word-list structures to examine the foundational question of whether and how group recall changes the collective memory organization because these materials offer many advantages at this initial stage. Like narratives, related stimuli such as categorized word lists or words with different emotional valence offer externally imposed conceptual structures, thereby providing a systematic link across these materials. These stimuli also allow the use of many measurement tools, providing initial quantitative insights into the emergence of collective memory. These foundational steps provide both a link to the existing literature on memory organization in individuals and offer a path to scale up to narratives in future investigations. In this vein, we have successfully followed this route to measuring collective memory content, first in studies using word stimuli and then scaling up to narratives. In Maswood et al. (2019), participants recalled a recent exam they took for which we measured the formation of collective memory following collaborative remembering compared with individual remembering. In this investigation, we used not only more complex and narrative productions but also memories that were autobiographically situated, amplifying the naturalistic progression toward understanding memory in everyday life and memories that have group-identity elements (all members of a group had taken the same exam). In this way, using laboratory methods, we could better understand the underlying cognitive principles and quantitatively characterize how these shared representations emerge.
Returning to memory structure alignment, past work has captured this by measuring organization at local levels, that is, clusters of items within the recall sequence that match across collaborators (Congleton & Rajaram, 2014; Greeley et al., 2024). In this article, we describe two quantitative tools we have recently applied to advance the study of synchronized recall organization at a global level. The first tool, consisting of representational similarity analysis (RSA; Kriegeskorte et al., 2008), allows us to go beyond local-level clusters by quantifying the full sequence of recall and the interrelated positioning of all elements in recall. This process captures memory organization at a global level, affording the holistic assessment of the emerging collective narratives. The second tool, network analysis, offers a different approach to viewing the global-level arrangement of recall of information across many people, thereby allowing a holistic assessment of collective knowledge.
Representational Similarity Analysis: Holistic Memory Organization
RSA is a quantitative tool that offers a global characterization of response patterns when people recall past information. Applied to free recall data, RSA captures retrieval patterns in a comprehensive fashion, making it possible to incorporate information about (a) what people remember, (b) what was not remembered, and (c) the relative position of all to-be-remembered material (Jin et al., 2025). This application extends the use of RSA beyond neuroimaging research (Kriegeskorte et al., 2008) and human/large language model similarity assessments (Ogg et al., 2025). As such, applying RSA to recall data provides a window into how memories are reconstructed at a global level, incorporating information about memory content and recall order to construct an interconnected matrix. This approach provides a view of how in this memory representation each element is connected to every other element and the position of each element within this overall memory representation.
While Jin et al. (2025) provided a comprehensive overview, in brief, the approach involves computing the pairwise recall distance between all recalled and nonrecalled material. As a first step, we have applied this analysis to data consisting of the recall of word lists (Jin et al., 2025). Consider the following example: “Apple” was recalled first (Output Position 1), “banana” was recalled second (Output Position 2), and “chair” was recalled much later (e.g., Output Position 12). Here, the recall distance between “apple” and “banana” is 1, whereas the distance between “apple” and “chair” is 11. A word that is not recalled is coded with an arbitrarily large number—one that is much larger than the possible set of stimuli—such as 999. With nonrecalled words coded this way, distances can be computed as before. Representing recall outputs in this fashion involves making three assumptions: All material could be recalled given a long enough timeline or the right cue, material output later during recall is “closer” to nonrecalled material, and all nonrecalled material is represented equally (e.g., 999 − 999 = 0). A recall matrix including all pairwise distances is referred to as a “representational dissimilarity matrix” (RDM). We leveraged this approach to examine how global retrieval patterns are synchronized by collaborative recall, considering full memory matrices (including all to-be-remembered material) and partial matrices (e.g., separating to-be-remembered material by emotional valence; Jin et al., 2025). As such, the value of this approach stems from how these matrices are used. Specifically, these matrices are well suited for a comparison (via correlation or through other means) that can highlight the similarity between representations both within and across individuals. Within-individuals comparisons (e.g., individual recall before and after collaboration) reveal the global-level changes in an individual’s recall organization, whereas across-individuals comparisons (e.g., in individual recall after collaboration) reveal whether collaboration has changed and made similar the global-level organization across people. Table 1 and Figure 1 provide examples of how this approach is implemented.
Experimental Methodology Highlighted by Jin et al. (2025)
Note: In the nominal (III) condition, participants completed three individual recall phases. In the collaborative (ICI) condition, participants completed two individual recall phases separated by a collaborative recall phase. This framework affords a well-controlled assessment of how collaborative recall changes not only the contents of memory but also the interconnections among them. Adapted from “Collaborative Recall Changes the Global Organization of Memory: A Representational Similarity Analysis of Social Influences on Individual and Collective Memory Organization,” by J. Jin, H.-Y. Choi, G. D. Greeley, N. W. Pepe, E. A. Kensinger, A. Mohanty, and S. Rajaram, 2025, Journal of Experimental Psychology: General, 154, 1761–1783. Reproduced with permission from American Psychological Association; and based on Choi (2015).
Using RSA to capture collective memory organization. RSA can be applied to free recall data by labeling all to-be-remembered material with either the observed recall serial position or 999 (Step 1; top). An RDM can be constructed by computing the distance between all unique item pairs and vectorizing this distance information (Step 2; top). Vectorized RDMs can be compared to other vectorized RDMs, within or between participants, to provide insight into retrieval strategy similarity at a holistic level (Step 3; top). Following these steps when approaching collaborative data stemming from the experiment described in Table 1 would result in seven RDM vectors for each three-person group (Group A; middle). Specifically, each individual has two individual RDM vectors (Recall 1 and 2), and each group has one group-recall RDM vector (Recall 2). With these RDM vectors prepared, MDS (Borg et al., 2013) can be leveraged to visualize the influence that a collaborative recall phase can have on individual retrieval strategies (MDS plot; bottom). A clear trend emerges using this plotting strategy; individuals tend to drift toward the group-level strategy, demonstrating not only a propensity to recall more of the same material (collective memory) but also a tendency to adopt common interconnections. RSA = representational similarity analysis; RDM = representational dissimilarity matrix; MDS = multiple dimensional scaling; ICI = individual-collaborative-individual. Adapted from “Collaborative Recall Changes the Global Organization of Memory: A Representational Similarity Analysis of Social Influences on Individual and Collective Memory Organization,” by J. Jin, H.-Y. Choi, G. D. Greeley, N. W. Pepe, E. A. Kensinger, A. Mohanty, and S. Rajaram, 2025, Journal of Experimental Psychology: General, 154, 1761–1783. Reproduced with permission from American Psychological Association.
Each RDM can be vectorized and, with recall distance vectors in hand, compared with one another via correlation or by considering their similarity using distance metrics. For example, computing the distance between pairs of recall distance vectors within a group (see Fig. 1) produces a square matrix of distances when visualized, affording a look into whether individual retrieval patterns drifted toward the group pattern. As shown in Figure 1, that pattern clearly indicates that collaboration pulled postcollaborative recall performance closer to the group’s performance. As to what about the recall pattern is becoming more similar, because recall distance vectors are very high dimensional, the distance between them is informative for summarizing general closeness or similarity but not for zeroing in on particular similarities. If that is a goal, a method that focuses more on particular item-to-item transitions (such as the network approach we describe below) could be more suitable.
The comprehensive nature of this RSA application lends itself to capturing complex patterns, including those featured in narratives, episodes, conversations, or stories that give rise to collective memory. An application such as RSA and other suitable measures such as hypergraphs or computational choices that capture changes to network structures over time (e.g., Peel & Clauset, 2014) can help quantify in a comprehensive fashion the magnitude of changes that occur in the recall output from one rendition to the next, creating a different story.
Network Analysis: Interconnections in Retrieval
The second tool of interest characterizes a corpus of recall data as a network, providing a window into the collective and more idiosyncratic ways that people retrieve information. Although network analysis is prevalent in the study of semantic memory (Kumar, 2021), we leverage this tool to quantify retrieval overlap across people to get at collective memory content and organization. This approach allows us to arrive at a qualitative (visual) and quantitative summary of local response patterns across many respondents and capture interconnections in memory at a collective level. Applied to free recall or fluency task data, such a network characterization reveals how people tend to transition between responses (Greeley et al., 2025). This provides a glimpse into memory search and a test of how stable response patterns are among members of a collective. Like the RSA application described above, the result is a tapestry with interconnected elements, although of a somewhat different sort. However, rather than taking a bird’s-eye view, this representation “zooms in” and considers only the most local information (e.g., Word A → B, B → H, H → M). Across a collective, this local information can be aggregated to reveal trends that can be summarized in network terms.
Greeley et al. (2025) provided a comprehensive overview of how network analysis can be applied to recall data to characterize collective retrieval patterns. Given raw recall data (i.e., what was recalled and in what order), constructing a network is relatively straightforward. For each person, each recalled unit is represented as a node in the network, with nodes being connected by an edge if the recalled units appeared in adjacent positions during recall. 1 Consider the following example recall protocol: apple, banana, desk, couch, mountain, lamp, shoes, pear. For this single recall, a network representation would consist of eight nodes (one for each word recalled) and seven edges (apple-banana, banana-desk, etc.), resulting in a chain. This is more interesting, and informative, when recall outputs from multiple people are combined. When a network is constructed on the basis of the responses of a collective, it is possible to discover common retrieval patterns, over and above common responses, thus providing information about not only the retrieved content but also the organization of this content.
In a recent study, we examined the stability of response patterns across time and in different testing contexts, harnessing data from five studies conducted between 2011 and 2021 (Greeley et al., 2025; Fig. 2). Departing from laboratory studies of collective memory, we operationalized collective memory as a shared body of knowledge (e.g., DeSoto & Roediger, 2019; Roediger & Abel, 2015). Specifically, in each study, participants were asked to recall as many U.S. cities as they could for a set period of time. Like national and cross-national surveys using individual-level responses to capture collective-level representations about historic figures (e.g., U.S. presidents; DeSoto & Roediger, 2019) or events (e.g., World War II; Abel et al., 2019), examining recall of U.S. cities in a fluency task can reveal patterns of semantic knowledge about how people collectively represent their geographic contexts. In our studies, participants always worked alone during this task, which was typically used as a distractor task within memory experiments and once as a stand-alone task.
Using network analysis to capture collective memory organization. Applying a network approach in this context, we assessed collective organization by constructing networks—cities were connected to other cities if they were recalled in adjacent positions (top). The more frequently cities appeared in adjacent positions across participants, the stronger their connection. The most popular connections overall are summarized in rank order (bottom). Such a network perspective provides insight that goes beyond a content-only analysis and suggests several robust patterns for how cities are recalled (e.g., geography-informed retrieval). From “Collective Memory and Fluency Tasks: Leveraging Network Analysis for a Richer Understanding of Collective Cognition,” by G. D. Greeley, T. Peña, N. W. Pepe, H.-Y. Choi, and S. Rajaram, 2025, Canadian Journal of Experimental Psychology, 79, 61–73. Reproduced with permission from American Psychological Association.
We found that the same cities were recalled at similar rates across projects conducted over a decade (2011–2021) and in different testing contexts (e.g., in-person and online). Moreover, the network analysis results suggested a high degree of similarity in how these cities were recalled, that is, what cities were recalled in adjacent positions. Looking into the retrieval process, the network depicts the most common city-to-city transitions. Several patterns were very robust, especially considering the number of potentially valid cities (and the many more possible response sequences). People tended to organize or cluster responses in two ways—prototypic and geographic. For example, New York City and Los Angeles are massive metropolitan areas that fit the prototype of “city” and were frequently clustered despite not being close geographically. At the same time, major cities within geographic regions also tend to be clustered; Austin, Dallas, and Houston were frequently reported in adjacent positions. Although major cities in their own right (based on population), it seems like “Texas cities” might be the organizing principle here.
The application of this network approach need not be restricted to word-list (or city-names) recall. Echoing the flexible nature of the RDM/RSA methodology described above, a network approach is well suited for capturing more nuanced patterns in narratives, stories, and beyond. So long as the material in question can be discretized, a network can be constructed, summarized, and compared with other networks. Likewise, representing narratives as networks to uncover richer dimensions than afforded by a content or occurrence-only analysis has precedent (Min & Park, 2019). We suggest that this approach could be useful when considering collective memory, particularly when the focus is on how collectives reconstruct their past. This approach could prove especially valuable in cases in which members of different large-scale collectives (e.g., sports fans, political parties, religious groups) may produce similar narratives at the level of content but emphasize and associate narrative elements in different ways. Such patterns could underscore collective overclaiming, in which members of a particular group (e.g., a nation) overestimate their own collective responsibility for collective successes. For example, survey respondents in the United States and United Kingdom estimated that their respective countries played a greater than 50% role in helping end World War II (Roediger & Zerr, 2022). In this way, on the one hand, an exploratory investigation could reveal general patterns characteristic of particular collectives, for example, as noted above for “massive cities” and “Texas Cities.” On the other hand, a confirmatory investigation could leverage a priori expectations about how members of different collectives build narratives.
Concluding Remarks
In this article, we described our novel applications of two prevalent quantitative tools in the scientific literature, RSA and network analysis, to study the holistic organization of collective memory. Going beyond an assessment of what collectives remember, these quantitative tools provide a window into the interconnections within this memory content and the way these interconnections become synchronized across group members. These novel applications offer three notable advances. First, their application provides a global or holistic view of how memory becomes organized both within and across individuals, offering an advantage over the capture of small units of organization in assessing the structure of collective memory. Second, considering collective organization over and above a content-only analysis also provides theoretical degrees of freedom. Whether the focus is on what organizational features make a collective similar (or what patterns result in stark differences), the tools we highlighted here are capable of capturing the most granular shifts as well as the bigger picture. Last, these quantitative tools have the methodological flexibility and richness to apply to a range of to-be-remembered information, making them suitable for assessing collective memory in a variety of contexts.
Recommended Reading
DeSoto, K. A., & Roediger, H. L., III. (2019). (See References). Describes research based on tasks that measure collective knowledge using individual-level responses to capture collective-level representations about historic figures.
Greeley, G. D., Peña, T., Pepe, N. W., Choi, H.-Y., & Rajaram, S. (2025). (See References). Leverages a fluency task and network analysis to capture collective knowledge about U.S. cities.
Jin, J., Choi, H.-Y., Greeley, G. D., Pepe, N. W., Kensinger, E. A., Mohanty, A., & Rajaram, S. (2025). (See References). Integrates experimental findings with the computational technique of representational similarity analysis to assess changes in the global-level organization of individual and collective memory following social remembering.
Rajaram, S., & Pereira-Pasarin, L. P. (2010). (See References). Explores cognitive research on collaborative remembering and collective memory and presents a conceptual framework for investigating the cognitive processes that shape these memory functions.
Weldon, M. S., & Bellinger, K. D. (1997). (See References). Shows the disruptive consequences of collaborative remembering on group performance and its benefits for later individual remembering.
