Mnemonic convergence in social networks: The emergent properties of cognition at a collective level

To explore whether the differences in clustering between the two conditions influenced the degree of mnemonic convergence that the communities reached, we conducted a mixed factorial ANOVA, with time (preconversation vs. postconversation) as a within-network variable and network type (clustered vs. nonclustered) as a between-network variable. This analysis was performed at the network level, with the mnemonic convergence score as a dependent variable. We found a significant main effect for time, [F_{(1, 12)} = 85.62, P < 0.001] and marginally significant effects for network type [F_{(1, 12)} = 4.27, P = 0.061] and for the interaction between time and network type [F_{(1, 12)} = 4.56, P = 0.054]. The degree of mnemonic convergence increased from preconversation to postconversation for both the clustered (P < 0.002) and the nonclustered (P < 0.001) conditions. As the interaction suggests, the increase was larger in the nonclustered condition than in the clustered condition (P = 0.054) (Fig. 3 A and B). These effects were in the hypothesized direction, but only marginally significant; thus, we conducted additional analyses to more precisely explore the effect of network structure on mnemonic convergence.

An important difference between the two network structures is the distance between the participants in the conversational network. The degree of separation, defined as the number of links in the shortest path between two nodes, ranges from 1 to 5 in the clustered condition and from 1 to 3 in the nonclustered condition. We posit that the smaller the degree of separation between any two participants, the more similar their mnemonic representations should become after network conversations. For instance, if participant P1 interacts with participant P3, and P3 interacts with P5, but P1 does not interact with P5, then the mnemonic similarity between P1 and P3 would be expected to be larger than the mnemonic similarity between P1 and P5. Such a prediction, if confirmed, could explain the difference in mnemonic convergence between the two network structures.

To test for a degree of separation effect, we first computed a mnemonic alignment score by subtracting the preconversational mnemonic similarity score from the postconversational mnemonic similarity score for each pair of participants (Eq. 3). This mnemonic alignment score, computed at a dyadic level, constituted the dependent variable for this analysis. We used separate linear mixed model analyses for the clustered and the nonclustered conditions, given the differing degrees of separation between the two conditions. The degree of separation, a between-subjects factor, constituted an independent variable and was nested by network. We found a significant effect for degree of separation in the clustered condition [F_{(1, 288.717)} = 10.27, P < 0.002], but not in the nonclustered condition [F_{(1, 74.936)} = 1.75, P = 0.19]. A polynomial contrast revealed that the effect was linear in the clustered condition (P < 0.006), but not in the nonclustered condition (P = 0.20) (Fig. 3 C and D). Despite this difference in the linear nature of the effect, there was no statistical difference between the two conditions in the pattern of mnemonic alignment when the analyses were restricted to 1–3 degrees of separation. In both conditions, only participants who were 1, 2, or 3 degrees of separation away from one another had alignment scores significantly larger than 0 (P < 0.001 for all three comparisons). These results provide support for the hypothesis that the conversational network structure impacts the alignment of participants’ memories. (An analysis that solidifies this conclusion is provided in SI Materials and Methods; see Fig. S1.) The question then becomes what are the processes by which conversations align the participants’ memories?

As discussed above, a burgeoning literature has shown that jointly remembering previously studied information makes mentioned information more likely to be remembered by the conversational partners in subsequent individual recollections (7, 12, 20). At the same time, unmentioned information related to what is discussed during the conversation is less likely than unmentioned and unrelated information to be subsequently remembered by the interactants. It follows, then, that the reinforcement and retrieval-induced forgetting effects triggered during conversational interactions could similarly affect the interactants’ memories, thereby impacting the emergence of collective memory.

To test for this possibility, we analyzed the content of all participants’ conversations. Following Cuc et al. (12), we first classified, for each conversation, all of the items of information that the participants studied as a function of their conversational retrieval practice status, as follows: (i) items mentioned during the conversation were labeled Rp+ (retrieval practice plus), (ii) items not mentioned during the conversation but related to the mentioned items, that is, those from the same category as the practiced ones, were labeled Rp− (retrieval practice minus), and (iii) items unmentioned during the conversation and unrelated to those mentioned were labeled Nrp (no retrieval practice). The labeling of items was relative to each of the three conversations in which a participant was involved during the conversational remembering phase, such that an item coded as Rp+ during the participant’s first conversation could be coded as Rp− during that participants’ subsequent conversation. We did not account for the source of the information during the conversation (i.e., who was the speaker and who was the listener), because previous research showed that during conversational recall tasks, the speakers and listeners experience similar degrees of reinforcement and retrieval-induced forgetting effects (12).

Using the foregoing labeling scheme, we computed reinforcement/suppression (R/S) scores for each of the 16 initially studied items for each participant. If an item was mentioned during a conversation, and was thus an Rp+ item, it received a (+1) score on the R/S scale. Similarly, if an item was not mentioned during a conversation but was related to the mentioned item (i.e., Rp−), it received a (−1) score on the R/S scale. Items unmentioned and unrelated to the mentioned items (i.e., Nrp) received a score of 0 on the R/S scale. The final R/S score for each participant was cumulated across the three conversations that he or she had in the network. For instance, if an item was mentioned in all three conversations that a participant had in the network, then that item was coded as having a (+3) R/S cumulative score, whereas if the item was part of the category mentioned during a participant’s conversations but itself was never mentioned in any of the three conversations, then that item was coded as having a (−3) R/S cumulative score. Importantly, each item was categorized based on its R/S score, which could range from −3 to +3. The scores computed using this procedure are designated as “mixed” R/S cumulative scores, because the status of an item could change from Rp− to Rp+ to Nrp in the different conversations in which participants were involved. Based on previous studies that used a selective practice paradigm, we predicted that items with positive R/S scores would experience a reinforcement effect, whereas items with negative R/S scores would experience a retrieval-induced forgetting effect. To capture these effects, for each item we computed a mnemonic difference score by subtracting the preconversational recall score from its postconversational recall score (Fig. 4 and Eq. 4 in Fig. 2). A positive value for this mnemonic difference score indicates a reinforcement effect, whereas a negative value indicates a retrieval-induced forgetting effect.

We conducted a mixed factorial ANOVA with R/S item type as a within-network variable with five levels (+2, +1, 0, −1, and −2) and network type (clustered vs. nonclustered) as a between-network independent variable. For this analysis, we aggregated individual-level scores into network-level scores to avoid missing-data issues derived from using a five-level repeated-measures variable. To do so, we computed mnemonic difference scores for each participant and then averaged these scores across all participants within each network, separately for each R/S item type. These network-level scores constituted the dependent variable for this analysis. Even though the R/S scores ranged from −3 to +3, we only ran analyses with a range between −2 and +2, owing to an insufficient number of data points for the extreme scores (less than 3% of data points had R/S scores of +3 and −3). We found a main effect for R/S item type [F_{(4, 9)} = 21.68, P < 0.001], but not for network type (P = 0.81) or for the interaction (P = 0.96). Furthermore, the effect was qualified by a linear trend, with mnemonic difference scores decreasing as R/S scores decreased [F_{(1, 12)} = 51.72, P < 0.001] (Fig. 5A, black line), showing the expected pattern.

To disentangle the effects of reinforcement and retrieval-induced forgetting on subsequent memories, we conducted the same analysis as before, but this time using a subset of the data for which we could compute “pure” R/S cumulative scores. To compute these scores, we calculated the mnemonic difference scores for only the items that could be unequivocally categorized as reinforced, suppressed, or baseline, based on each participant’s three conversations. Thus, if an item constituted an Rp+ item in two of the three conversations of a participant and an Nrp item in the remaining conversation, then the item was twice reinforced and experienced no suppression, for a pure R/S cumulative score of +2. Similarly, if an item constituted an Rp− item in one conversation of a participant and an Nrp in the other two conversations of that participant, then the item was suppressed once and experienced no reinforcement—a pure R/S cumulative score of −1. Items that were categorized as Nrp items for all three conversations of a participant constituted baseline items and were assigned a pure R/S score of 0.

Because the analysis was conducted at a network level, we averaged the mnemonic difference scores across the 10 participants in the network, separately for the different types of R/S items. Importantly, 75% of the data points entered into the mixed R/S scores analysis were categorized as pure R/S instances. As in the above analysis, we found a main effect for R/S item type [F_{(4, 6)} = 38.19, P < 0.001], but not for network type (P = 0.51) or for the interaction (P = 0.61). Furthermore, the effect was qualified by a linear trend, with mnemonic difference scores decreasing as the pure R/S scores decreased [F_{(1, 9)} = 70.95, P < 0.001] (Fig. 5A, red line). Items that were twice or once suppressed had a significantly lower mnemonic difference score compared with the baseline items (P < 0.05). Items that were twice reinforced had a significantly higher mnemonic difference score than the baseline items (P < 0.03), unlike once-reinforced items (P = 0.35) (Fig. 5A, red line). Of note, the analyses involving pure R/S scores were performed for 11 of the 14 networks, because three networks had no items that could be designated pure Nrp items.

We have provided evidence that cognitive mechanisms triggered during conversational remembering (i.e., reinforcement and retrieval-induced forgetting effects) influence the participants’ subsequent memories. But do the sociocognitive processes triggered during conversations affect mnemonic convergence at a network level? To answer this question, we followed Weldon (23), who pioneered the use of social network methodology to explore the formation of collective memories. According to this approach, collective memories could be represented as the aggregation of individual memories of the members who belong to the community (i.e., a participants-by-memory items matrix). By transposing this matrix into an item-by-item matrix containing the number of participants that remembered the items in common, and then dichotomizing the matrix using a majority rule (n > 5 for a 10-member network), we obtain a binary matrix that captures the community’s collective memory. In a visual representation of such a network, nodes are items, and connections indicate that the members’ linked nodes (items) are remembered by a majority of the community (in our case n > 5; Fig. 2, last row).

We computed such binary matrices separately for each network, for both preconversational and postconversational recalls. Based on these binary matrices, we calculated item-level closeness centrality scores for each of the 16 items (21). Defined as the graph-theoretic distance of a given node to all other nodes, this measure captures how central each of the 16 items is to the collective memory of a community. We used a normalized measure for closeness centrality (21), such that high closeness centrality scores indicate higher centrality in the network’s collective memory (Fig. 2, Eq. 5). We then subtracted the preconversation closeness centrality from the postconversation closeness centrality to capture item-level closeness centrality dynamics from preconversation to postconversation.

For a measure of conversationally triggered reinforcement and retrieval-induced forgetting effects, we used the mixed R/S scores as described above. We computed item-level mixed R/S cumulative scores by averaging the R/S scores over all of the 15 conversations that took place in the conversational recall phase for each network. To investigate whether the R/S scores predict closeness centrality difference scores, we ran separate regression analyses for each of the 14 networks. Both the R/S scores and the closeness centrality difference scores were computed at an item level for each of the 16 items that participants studied. We did not expect to find differences between the two conditions in the degree to which an item’s R/S score predicts its closeness centrality difference score. To control for item memorability effects, we included the item’s preconversational recall score in the regression analysis. For 11 of the 14 networks (six clustered and five nonclustered), the R/S scores were significant predictors of closeness centrality difference scores (P < 0.05); for two networks, the R/S scores were marginally significant (P = 0.08 and 0.11); and for one network, the R/S score was not a significant predictor (P = 0.47).

To provide further support linking the conversationally triggered sociocognitive processes with the network level effects, we conducted a more fine-grained analysis. As described above, the average R/S scores were computed at a network level, with a single cumulative R/S score corresponding to each of the 16 items studied by the participants. Based on these R/S scores, for each network we categorized the 16 items into quartiles, from highest to lowest R/S scores. As before, for each of the 16 items, we subtracted the preconversation closeness centrality from the postconversation closeness centrality. This difference captured item-level closeness centrality dynamics from preconversation to postconversation. We then averaged these centrality difference scores separately for the four R/S quartiles. Using these network level closeness centrality difference scores as a dependent variable, we conducted a mixed factorial ANOVA with R/S quartiles as a within-network variable and network type (clustered vs. nonclustered) as a between-network independent variable. We only found a main effect of R/S quartiles [F_{(3, 10)} = 70.43, P < 0.001], qualified by a linear trend [F_{(1, 12)} = 160.20, P < 0.001], as expected (Fig. 5B). Taken together, these two latter analyses indicate that the emerging collective memory is shaped by the sociocognitive processes triggered during conversational remembering. Items that become central to the collective memory of a community are those that are mentioned frequently and thus are reinforced in conversations among the members of the community. In contrast, the items that become peripheral to the collective memory of a community are those that are suppressed during conversational remembering.