Neural pattern stability within events is similar in young and older adults

Abstract

Understanding how the brain segments continuous experiences into discrete events is essential for uncovering the mechanisms of episodic memory. Here, we investigated whether stable neural representations are maintained within unfolding events, how these representations differ between young and older adults, and whether they predict later memory. Young (N = 42) and older adults (N = 42) watched a 35-min clip from BBC’s Sherlock while EEG was recorded. Using representational similarity analysis, we quantified neural pattern similarity within events and across event boundaries. Neural similarity was significantly higher within events than across events, even after controlling for perceptual changes in the stimulus, indicating stable neural representations during ongoing experience. This within-event neural stability was observed in both young and older adults, although evidence was stronger in young adults. Despite robust neural effects, neither within- nor across-event neural similarity significantly predicted subsequent free recall at the event or subject level. These findings suggest that neural stability reflects top-down maintenance of coherent event representations during perception, rather than directly determining later recall success. Importantly, both age groups showed greater within-event than across-event neural similarity, but follow-up analyses did not provide sufficient evidence to conclude equivalence between younger and older adults.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-59124-3.

Introduction

Although everyday life unfolds as a continuous stream of information, we experience it as a series of events separated by event boundaries, segmenting it into meaningful units^1–3. We maintain active representations, or event models, that remain stable during an ongoing event but need to be updated when one event ends and a new one begins, as the current model is no longer accurate^4,5.

Event boundaries play a critical role in structuring long-term memory by parsing continuous experience into discrete, meaningful events^3,5. Individuals who more reliably detect event boundaries also tend to show superior memory for those events^6–8, even after accounting for other cognitive factors associated with memory performance⁹. Event boundaries are thought to reflect moments at which ongoing event models are updated in response to prediction error or contextual change^5,10. These updating moments are also associated with rapid memory reinstatement of the just-encoded content¹¹.

Although event boundary perception and the updating it triggers clearly contribute to memory, a complementary account emphasizes the role of stable neural representations within ongoing events. Maintaining a coherent event model across time may promote the integration of information encountered at different moments, yielding a more unified and durable memory. This idea aligns with classic work showing that constructing and sustaining a coherent mental model supports comprehension of unfolding experience^12,13, whereas failures to establish such global coherence lead to poorer memory¹⁴.

What remains less clear is whether stability within an unfolding event contributes to episodic memory beyond supporting moment-to-moment comprehension. Many accounts of event memory emphasize boundary-related mechanisms—such as event model updating or hippocampal reactivation at event offsets—as the primary drivers of long-term memory encoding. From this perspective, stable neural activity during an event may play only a limited or indirect role. Thus, whether within-event neural stability makes a substantive contribution to episodic memory, above and beyond boundary-related processes, remains an open empirical question.

This question is particularly important in the context of aging. Because event segmentation helps organize memory, age-related changes in segmentation have been proposed as one source of age-related memory decline. In fact, when asked to explicitly identify event boundaries, older adults tend to segment more idiosyncratically compared to young adults^9,15, especially when prior knowledge is limited¹⁶. However, studies employing more covert measures of event processing—including eye movements^17,18, neural activity¹⁹, and reading times^20,21 have yielded mixed results.

Thus, age-related memory declines may not be due solely to changes in segmentation and updating processes but could also be due, in part, to older adults’ difficulty maintaining coherent event models over time, especially because event model maintenance requires sustained attentional control and working memory cognitive resources that often decline with age^22,23. Individuals actively maintain coherent mental models of ongoing events by integrating perceptual input, recent experiences, and semantic knowledge to support comprehension and memory^13,24,25. Critically, these models persist beyond immediate perception, allowing recently encountered information to remain active even when it is no longer present in the environment²⁶.

Neurophysiological research demonstrates that stable patterns of neural activity emerge during unfolding events and shift at event boundaries^27–29. Using approaches such as representational pattern similarity, multivariate state analysis, and temporal-dynamics modeling, converging evidence shows transient changes in brain activity at event boundaries^30–32 and greater neural pattern similarity within events than across boundaries³³. Critically, these neural mechanisms may vary across the lifespan. Despite age-related differences in behavioral segmentation, Kurby and Zacks found no age-related differences in neural responses at event boundaries¹⁹. To date, however, age-related changes in event model maintenance have not been directly examined.

In the current study, we investigated the stability of event models during naturalistic movie viewing in young and older adults and its relationship with episodic memory performance using EEG pattern similarity analyses. We first tested whether neural pattern similarity was greater within events than across adjacent events, reflecting stable event representations. We then examined whether within-event neural stability differs between young and older adults, testing whether this effect changes with age. Finally, we tested whether within-event neural stability predicts subsequent memory performance, or whether episodic memory is more strongly linked to processes occurring at event boundaries.

Results

First, we examined EEG pattern similarity to evaluate whether it differed within and across events in both young and older adults. Next, we assessed free recall performance and tested whether EEG pattern similarity predicted memory performance in young and older adults.

Age-related differences in EEG pattern similarity

To examine whether EEG pattern similarity differed within versus across events and whether this effect varied by age, we conducted a linear mixed-effects analysis predicting trial-level EEG pattern similarity values. Z-scored perceptual change was included as a covariate to control for low-level visual variability. Two nested models were compared using maximum likelihood estimation: a baseline model with fixed effects of Trial Type, Age Group, and their interaction, and a second model adding perceptual change. Both models included random intercepts for participants and events. Including perceptual change significantly improved model fit (baseline model: AIC = – 4575.10; second model: AIC = – 5128.30; ΔAIC = 553.20), as confirmed by a likelihood ratio test, χ²(1) = 555.25, p < 0.001. Given the improvement without overfitting, this model was selected for interpretation.

In the final model, EEG pattern similarity measured using representational similarity analysis (RSA) was significantly higher within events than across events (β = 0.012, p = 0.034; see Fig. 1), indicating greater neural stability during continuous experiences. Perceptual change was a strong negative predictor of EEG similarity (β = –0.089, p < 0.001), indicating that increased low-level visual variability was associated with reduced neural similarity. To further examine whether perceptual change differentially influenced EEG similarity in young and older adults, we tested an additional model including the Perceptual Change × Age Group interaction. This interaction was not significant, β = 0.001, SE = 0.002, t = 0.51, p = 0.610, and including it did not improve model fit relative to the more parsimonious model, χ²(1) = 0.26, p = 0.610; AIC =  − 5128.35 for the base model vs. − 5126.61 for the interaction model. Thus, perceptual change was a strong negative predictor of similarity, but its effect did not reliably differ between young and older adults. However, perceptual change did not differ significantly across trial types (M_within = 0.048, SD = 0.021; M_across = 0.051, SD = 0.022, t(33) = –0.64, p = 0.530, 95% CI [–0.0113, 0.0059]), suggesting that trial-type differences in neural similarity are not explained by perceptual variability.

Fig. 1.

Model-adjusted EEG pattern similarity within and across events by age group. Violin plots display EEG pattern similarity values for young and older adults, derived from a linear mixed-effects model including Trial Type, Age Group, their interaction, and perceptual change as fixed effects, with random intercepts for participants and events. The model revealed a significant main effect of trial type, with higher similarity within events compared to across events (p < 0.05). Perceptual change was also a significant predictor of similarity (p < 0.01). No significant main effect of age group or interaction effects were found (ps > .19). Diamonds indicate group means; dots represent individual data points; error bars show standard errors.

There was no significant main effect of Age Group (β = 0.008, p = 0.431), and no significant interaction between Trial Type and Age Group (β = 0.009, p = 0.278), indicating no reliable evidence that the within-event advantage in neural similarity differed by age group. However, Bayesian analyses provided a more nuanced characterization of these effects. Bayes factors were computed using the BayesFactor package in R³⁴. There was only anecdotal evidence supporting the null hypothesis of no overall Age Group difference in similarity (BF₀₁ = 2.13) and no Trial Type × Age Group interaction (BF₀₁ = 2.39).

At the group level, Bayesian results indicated strong evidence that young adults showed greater within-event than across-event similarity (BF₁₀ = 35.63), consistent with stable event representations over time. In contrast, older adults showed only anecdotal evidence for this effect (BF₁₀ = 2.04). Together, these findings suggest that the non-significant age-related effects in the frequentist models should be interpreted cautiously, as the Bayesian results indicate differing levels of evidence for within-event similarity across age groups rather than clear support for the absence of age differences.

To test whether the within- versus across-event similarity effect differed by age group, we computed a participant-level difference score by subtracting mean across-event similarity from mean within-event similarity. Positive scores indicated greater within-event similarity. The difference score was numerically larger in young adults (M = 0.02, SD = 0.04) than older adults (M = 0.01, SD = 0.03), but this difference was not significant, Welch’s t(80.69) = 1.18, p = 0.242, 95% CI [− 0.0060, 0.0234]. A Bayesian independent-samples t-test yielded BF₁₀ = 0.42 and BF₀₁ = 2.40, indicating modest evidence against an age-group difference. Because non-significance does not establish equivalence, we also conducted an exploratory equivalence test using raw bounds of ± 0.01 similarity units. The test was not significant because the upper-bound test did not reach significance, t(80.69) = − 0.17, p = 0.432. Thus, although both groups showed positive within–across difference scores, these analyses do not establish statistical equivalence between age groups.

Thus, the consistent finding is that within-event neural stability is greater than across-event neural stability. However, it should be noted that the temporal distance (lag) between the two segments of EEG data used in the within- and across-event trials was different. The across-event lag was always held constant at 6 s, whereas the within-event lag ranged from 4.6 to 205.76 s (mean = 43.74) due to variability in event durations. To address this, we conducted two additional analyses. First, we tested whether within-event neural similarity was influenced by the temporal distance between segments. A linear mixed-effects model revealed a small but significant main effect of lag: the further apart segments were, the lower their neural similarity (β = − 0.018, p < 0.001). Importantly, there was no significant interaction between lag and age group (p > 0.10). Age-group-specific correlations are reported in Supplementary Table S1.

Second, we conducted a lag-matched control analysis using similarity derived from the middle portions of each event. The temporal distance between the two segments was held constant for both within-event and across-event trials, ensuring that any differences in similarity could not be attributed to systematic differences in temporal separation. We identified the temporal midpoint of each event and extracted two 3 s segments: one from 3 to 6 s before the midpoint and one from 3 to 6 s after the midpoint. This model included Trial Type, Age Group, and perceptual change as predictors, with random intercepts for subjects and events. Results mirrored the main findings: lag-matched within-event similarity was significantly higher than across-event similarity, perceptual change was a strong negative predictor of similarity, and neither Age Group nor the Trial Type × Age Group interaction was significant. Full results are reported in Supplementary Table  S2.

Free recall performance

To examine age-related differences in participants’ ability to recall events from the movie, we conducted a generalized linear mixed-effects model (GLMM) with free recall accuracy (correct = 1, incorrect = 0) as the dependent variable, age group (young, older) as a fixed effect, and random intercepts for SubjectID and Event. The analysis revealed no significant main effect of age group (β = –0.076, SE = 0.218, z = – 0.35, p = 0.729), indicating that younger and older adults did not significantly differ in overall recall accuracy. Overall, younger adults’ mean recall was 0.352 (SD = 0.48), while older adults’ mean recall was 0.339 (SD = 0.47). A Bayesian independent-samples t-test on subject-level free recall proportions provided moderate evidence for the null hypothesis of no age-group difference, BF₀₁ = 4.06. Thus, the non-significant frequentist result was accompanied by moderate evidence favoring the absence of an overall age-group difference in free recall. These mean values correspond to accurately recalling approximately 12 of the 35 events. See Fig. 2 for the free recall values predicted from the model.

Fig. 2.

Percentage of recalled events by age group. Violin plots display the proportion of recalled events for young and older adults. A generalized linear mixed-effects model revealed no significant main effect of age group on recall accuracy (p = 0.729), indicating comparable performance. Larger filled points indicate group means; smaller dots represent individual data points; boxplots reflect medians and interquartile ranges.

Exploring event-level EEG pattern similarity as a predictor of free recall performance

To evaluate whether neural similarity predicted event memory, we analyzed the data at both event and subject levels. At the event level, we tested whether neural similarity predicted recall of the same event. Separate GLMMs were run for within- and across-event trials, with similarity, Age Group, and their interaction as fixed effects and random intercepts for subject and event. Adding z-scored perceptual change did not improve fit for either model, within-event: χ²(1) = 1.47, p = 0.225; across-event: χ²(1) = 1.78, p = 0.182, and AIC values were slightly lower for the baseline models. Therefore, we report the more parsimonious baseline models, noting that results were unchanged when perceptual change was included.

Within-event analysis

The baseline model revealed neither the main effect of within-event similarity on free recall accuracy (β = –0.111, SE = 0.369, z = – 0.30, p = 0.764), the main effect of age (β = – 0.081, SE = 0.220, z = – 0.37, p = 0.710), nor the Similarity x Age group interaction (β = 0.086, SE = 0.418, z = 0.210, p = 0.837) significantly predicted free recall performance. See Fig. 3a for predicted values from this model. A Bayesian subject-level comparison of within-event similarity for recalled versus unrecalled events provided moderate evidence for the null hypothesis of no overall recalled-versus-unrecalled similarity difference, BF₀₁ = 7.99. The same analysis also provided moderate evidence for no age-group difference in this effect, BF₀₁ = 4.25.

Fig. 3.

Model-predicted free recall accuracy by EEG pattern similarity, separated by age group. Plots show predicted probability of free recall as a function of mean similarity values, displayed separately for (a) within-event similarity and (b) across-event similarity. Shaded ribbons represent 95% confidence intervals.

Across-event analysis

For across-event trials, the baseline model also showed no significant effect of similarity (β = 0.108, SE = 0.467, z = 0.23, p = 0.820), age group (β = –0.072, SE = 0.226, z = – 0.32, p = 0.75) or similarity x age group interaction (β = 0.893, SE = 0.710, z = 1.26, p = 0.21) on free recall performance. See Fig. 3b. A Bayesian subject-level comparison of across-event similarity for recalled versus unrecalled events yielded moderate evidence for a recalled-versus-unrecalled similarity difference, BF₁₀ = 4.48. However, because this Bayesian comparison was conducted at the subject level rather than in the event-level mixed model, we interpret this result cautiously. The same analysis provided moderate evidence for the null hypothesis of no age-group difference in the recalled-versus-unrecalled similarity effect, BF₀₁ = 3.52.

These findings indicate that, at the level of individual events, neither within-event nor across-event neural similarity significantly predicted the likelihood of an event being recalled in either age group.

Next, we tested whether relative differentiation between within- and across-event neural patterns predicted memory for adjacent events. This analysis also addressed whether reduced differentiation may reflect poor boundary detection, blended event representations, and lower recall. For each boundary, differentiation was computed by subtracting across-event similarity from the mean within-event similarity of the two adjacent events. Recall was measured as accuracy for the events before and after the boundary, producing a combined score from 0 to 2. We then used a binomial mixed-effects model predicting recalled events from Boundary Differentiation, Age Group, and their interaction, with random intercepts for participant and boundary. The model showed no evidence of overdispersion, dispersion ratio = 1.064, p = 0.128. The corrected model revealed a significant Boundary Differentiation × Age Group interaction, β =  − 0.335, SE = 0.079, z = − 4.244, p < 0.001, OR = 0.72 (see Fig. 4). Among young adults, greater boundary differentiation showed a positive but non-significant trend toward greater combined recall, β = 0.110, SE = 0.059, z = 1.852, p = 0.064, OR = 1.12. Among older adults, greater boundary differentiation was associated with lower combined recall, β = − 0.225, SE = 0.067, z =  − 3.351, p < 0.001, OR = 0.80. This interaction remained significant when boundary-level perceptual change was included as a covariate, β = − 0.355, SE = 0.080, z =  − 4.455, p < 0.001, OR = 0.70, whereas perceptual change itself was not significant, β = 0.015, SE = 0.138, z = 0.106, p = 0.916. A BIC-approximated Bayesian model comparison provided evidence against a uniform main effect of boundary differentiation, BF10 = 0.023, BF01 = 42.68, but favored the interaction model over the main-effects model, BF10 = 118.91. Evidence favoring the interaction model over the age-only/null model was weaker, BF10 = 2.79. Thus, boundary differentiation did not uniformly predict better memory across participants; rather, its association with adjacent-event recall differed by age group.

Fig. 4.

Boundary differentiation and adjacent-event recall by age group. Predicted combined recall of the two events surrounding each boundary as a function of similarity-based boundary differentiation, separately for young and older adults. Lines represent model-predicted expected recall from the corrected binomial mixed-effects model with an observation-level random effect. Shaded regions indicate 95% confidence intervals. Combined recall scores ranged from 0 to 2, where 0 indicates recalling neither adjacent event, 1 indicates recalling one adjacent event, and 2 indicates recalling both adjacent events.

Subject-level analysis of EEG pattern similarity and event recall

Following the event-level analysis, we conducted a subject-level analysis to test whether individual differences in neural pattern similarity predicted overall memory performance. Participants’ total events recalled were modeled using two Negative binomial regression, appropriate for overdispersed count data, separately for mean within-event and across-event trials, with similarity, Age Group, and their interaction as fixed effects, and a random intercept for subjects. Consistent with the event-level results, in both models, neither main effects nor interactions significantly predicted free recall (see Table 1 and Fig. 5 for model output and predicted values). Complementary Bayesian model comparisons supported the frequentist results. Mean within-event similarity did not improve prediction of total recall beyond Age Group, BF₀₁ = 4.35, nor did mean across-event similarity, BF₀₁ = 3.05. A model including both similarity predictors also did not improve fit over an age-only model, BF₀₁ = 6.94, and adding Similarity × Age Group interactions did not improve fit over the main-effects model, BF₀₁ = 4.54. Thus, the non-significant subject-level similarity–memory associations were accompanied by moderate evidence favoring the null. The same subject-level analyses for cued recall, recognition memory, and temporal order memory showed the same pattern and are reported in the Supplementary Materials.

Table 1.

Negative binomial regression results for predicting number of events recalled.

Parameter	B	SE	z	p
Intercept	2.55	0.12	21.37	< 0.001
Mwithin	1.01	2.57	0.40	0.69
Age group (young vs. old)	− 0.06	0.17	− 0.37	0.71
Macross	− 3.99	3.69	− 1.08	0.28
Mwithin × age group	0.83	3.56	0.23	0.82
Macross × age group	1.62	4.78	0.34	0.74

Results from a negative binomial regression model predicting the number of events recalled. B = regression coefficient; SE = standard error. Mwithin = mean within-event similarity; Macross = mean across-event similarity. Age Group contrast is coded as Young vs. Old (0 = Old, 1 = Young).

Fig. 5.

Model-predicted number of events recalled by similarity, separated by age group. Plots show the predicted number of events recalled as a function of mean similarity values, displayed separately for (a) within-event similarity and (b) across-event similarity. Shaded ribbons represent 95% confidence intervals.

Discussion

The present study investigated age-related differences in neural pattern similarity during the encoding of dynamic events and their relationship to episodic memory. Using EEG pattern similarity, we examined whether neural stability differed within versus across events in young and older adults and whether these neural markers predicted free recall performance. Consistent with prior research, neural similarity was significantly higher within events than across events³³, supporting the idea that event models remain stable within continuous experiences^27,35.

This pattern of results aligns with theories of event cognition, which propose that the brain actively maintains internal representations of ongoing activity and resets these models at meaningful transitions in the narrative structure^2,25. According to these theories, event boundaries occur when predictions fail² or salient contextual shifts are detected^10,36,37, prompting the brain to update the current event model⁵. The reduced neural similarity across event boundaries in our EEG data likely reflects this updating process, wherein a previously stable model is no longer relevant and, thus, is replaced by a new model that represents the new situation. In contrast, the higher neural similarity observed within events reflects representational stability over time, consistent with the maintenance of a coherent event model during continuous experience³⁸.

One potential concern is whether within-event neural similarity reflects active maintenance over time rather than other encoding-related processes. However, several findings support interpreting this measure as event model maintenance. First, neural similarity was higher within events than across events, even after controlling for low-level perceptual features that vary moment to moment. Second, this within-event advantage occurred despite the larger temporal lag between within-event segment pairs (average lag ≈ 44 s) than across-event pairs (lag = 6 s). Third, neural similarity showed a modest decrease with increasing lag, consistent with the idea that longer events place greater demands on working memory and may lead to representational decay. Finally, the same pattern emerged in a lag-matched control analysis using mid-event segments. This reduces the possibility that the within-event advantage reflects boundary-related processes, such as event model construction or schema activation, because the segments were selected away from boundaries.

These EEG results align with findings from fMRI studies showing phasic increases in cortical^11,39,40 and hippocampal responses at boundaries^{27,35,41–43}. Importantly, event boundaries have also been shown to trigger rapid reinstatement of the just-encoded event, facilitating its consolidation into long-term memory^11,27,33,44. Our findings build on this literature by demonstrating temporally resolved EEG evidence that the brain maintains stable neural representations over the course of extended, dynamic events, and that these representations shift at points of transition between events. This shift is reflected in a reliable drop in neural pattern similarity across event boundaries, suggesting that the brain updates its event model when a meaningful change occurs.

It is also important to note that perceptual change significantly influenced neural similarity, with greater visual change being associated with lower neural similarity. However, perceptual change did not differ between within- and across-event trials, and within-event similarity remained significantly higher even after controlling for these low-level visual differences. Further, perceptual change did not predict memory performance. Thus, the brain still showed greater stability within events than across them, even after controlling for low-level perceptual fluctuations. Together, these results indicate that neural stability within events is not solely driven by bottom-up perceptual continuity. Rather, neural stability likely also reflects top-down processes—such as the maintenance of coherent event models—that support event segmentation and memory organization. From a theoretical standpoint, this raises a key question about how such top-down maintenance processes change with age. One account predicts that event model maintenance should decline in older adulthood due to increased demands on attentional control and working memory. In contrast, an alternative account predicts that event model maintenance is largely preserved in healthy aging, particularly when assessed using implicit neural measures rather than explicit task performance.

Consistent with prior work suggesting that some aspects of event processing may remain relatively preserved in healthy aging^19,45, we observed a within-event advantage in neural similarity in both young and older adults. However, follow-up analyses qualified the strength of this conclusion. Although the Age Group × Trial Type interaction was not significant, participant-level comparisons, equivalence tests, and Bayesian analyses did not provide sufficient evidence that the within-event advantage was statistically equivalent across age groups. Thus, the findings indicate the same directional pattern in both groups, but not strong evidence for equivalence. This pattern—reduced strength of evidence in older adults without a reliable age-group difference—parallels findings of increased neural idiosyncrasy and reduced intersubject synchrony in older adults during naturalistic viewing⁴⁶. These findings suggest that older adults retain evidence of within-event neural stability, although the present data do not support a strong claim that this mechanism is fully preserved to the same extent as in younger adults. Overall, the present findings add to a growing body of evidence indicating that key components of event cognition—including segmentation, updating, and potentially event model maintenance—remain largely intact in healthy aging, although the current evidence is insufficient to conclude that these processes are fully preserved relative to younger adults.

An unexpected finding was that neither within- nor across-event neural pattern similarity significantly predicted free recall at either the event or subject level. Complementary Bayesian analyses generally provided moderate evidence for the absence of several similarity-memory associations, particularly at the subject level. However, one subject-level Bayesian comparison provided moderate evidence for a recalled-versus-unrecalled difference in across-event similarity, which we interpret cautiously because this effect was not supported by the corresponding event-level mixed model. Prior research has shown that event boundaries trigger rapid reinstatement of just-encoded content, supporting long-term memory formation^11,33. Although we hypothesized that greater within-event stability and across-event distinctiveness (i.e., lower similarity) would support memory, event-level models provided no significant evidence for these effects.

In a follow-up boundary-level analysis, we tested whether relative differentiation between within- and across-event neural patterns predicted memory for adjacent events. The analysis revealed a significant Boundary Differentiation × Age Group interaction. Boundary differentiation was positively but non-significantly related to adjacent-event recall in young adults, consistent with the idea that well-defined event boundaries support memory⁸. However, the negative relationship between differentiation and memory in older adults is surprising. These results suggest that same neural pattern (boundary differentiation) may reflect different underlying processes in younger versus older adults. Memory for dynamic events depends on both the separation (segmentation) of events and the integration of information across events. Segmentation helps delineate where one event ends and another begins, whereas integration supports linking adjacent events into a coherent sequence.

In younger adults, these processes may operate in a balanced manner—events are sufficiently distinct to be separated yet still integrated enough to support recall. In contrast, the negative association observed in older adults may suggest that stronger segmentation may come at the expense of integration. For them, greater boundary differentiation may not indicate more effective event processing, but rather a shift toward separation without sufficient integration, which could impair memory.

At the subject level, trends suggest a potential relationship between neural stability or distinctiveness and memory, but these effects were not significant (see Fig. 5).

One potential explanation for these null results is that event model maintenance primarily supports online comprehension and the organization of experience, rather than directly determining recall success, which also depends on downstream retrieval processes. In contrast, event segmentation and model updating—marked by abrupt neural state changes at event boundaries—have been reliably linked to memory success, likely because these moments trigger encoding, consolidation, or hippocampal reinstatement processes^2,27,30. From this perspective, relative neural stability within events may reflect sustained representational continuity that supports understanding but is not itself sufficient to determine which event details are later recalled. Instead, segmentation-related dynamics may provide critical encoding signals, while factors such as attentional engagement, narrative coherence, and integration across events govern the likelihood that specific details are retained. This interpretation aligns with prior theoretical models that distinguish between within-event binding and across-event integration processes^37,47. Relatedly, a second possibility is that neural stability may be important for memory; however, variability in retrieval processes may override the influence of neural stability that occurs at encoding. A third possibility is that our binary measure of recall performance (remembered versus forgotten) may lack the sensitivity needed to detect the effects that neural stability exerts on episodic memory.

Several limitations should be noted. First, our measure of within-event neural stability—pattern similarity between two segments within each event—captured maintenance over time but not finer-grained dynamics within events. Although we showed that within-event similarity decreased with increasing temporal lag, future work could examine multiple time-resolved segments to better characterize how event representations evolve and distinguish sustained maintenance from gradual drift. Second, participants typed their recall responses, whereas prior studies often used spoken descriptions^27,33,48. Typing may impose additional cognitive or motor demands or reduce motivation compared to speaking. Future studies could compare typed and spoken recall to assess modality effects on memory performance. Third, participant demographics may have influenced performance. Most older adults were Caucasian lifelong U.S. residents, whereas the younger sample was more diverse, including some non-native English speakers or individuals less familiar with the cultural context of Sherlock. The show’s fast-paced dialogue and British accents may have posed comprehension challenges, reducing recall for some younger participants. Additionally, survey feedback indicated that some participants were uncomfortable with violent or suicide-related content, potentially lowering engagement and memory performance.

This study contributes to the growing body of research examining how the brain encodes, segments, and remembers continuous experiences. We show that both younger and older adults exhibit greater neural similarity within events than across event boundaries, consistent with the maintenance and updating of event models, respectively. However, our measure of neural stability during encoding did not robustly predict long-term episodic memory. These findings highlight the importance of considering both neural stability and differentiation when modeling event cognition processes across the adult lifespan.

Methods

Participants

A total of 88 participants were initially enrolled in this study: 45 young adults (ages 18–35; M = 26, SD = 4.35) and 43 older adults (ages 60–85; M = 72, SD = 4.45). A pilot study (n = 9 young adults) determined the effect size of the EEG pattern similarity for within vs. across event trials was Cohen’s d = 0.845. To be conservative, we assumed this effect would be half as large in older adults due to greater variability. Power analyses conducted in G*Power⁴⁹ with a two-tailed test, α = 0.05, power = 0.80, and d = 0.42 indicated that 36 participants per group were required to detect the within-subject effect.

Young adults were recruited from the Kansas State Cognitive and Neurobiological Approaches to Plasticity (CNAP) Sona system, whereas older adults were recruited from the Manhattan community using databases maintained by the Memory & Aging Lab. Participants were cognitively healthy, native English speakers or proficient in English, and had normal or corrected-to-normal vision (see Table 2 for demographics). The study protocol was approved by the Kansas State University Institutional Review Board (IRB; No. 11232). All procedures were conducted in accordance with relevant institutional guidelines and regulations and the Declaration of Helsinki. Written informed consent was obtained from all participants prior to participation, and all participants received $50 in compensation.

Table 2.

Demographic characteristics of participants.

Group	Gender (male/female)	Age range (years)	Mean age (M, SD)	Years of education (M, SD)	Highest level of education (%)
Young adults (n = 42)	17/25	18–35	25.78 (SD = 4.35)	18.1 (SD = 2.14)	High school (0), some college (17), bachelor’s (43), master’s (31), doctorate (10)
Older adults (n = 42)	14/28	60–85	71.64 (SD = 4.45)	17.7 (SD = 2.32)	High school (10), some college (8), bachelor’s (35), master’s (35), doctorate (12)

Values are reported as mean (SD). Percentages for education levels are calculated within each group and rounded to whole numbers.

Exclusion criteria

Participants were excluded if they had a history of neurological disorders (e.g., epilepsy, stroke, multiple sclerosis, traumatic brain injury) or were taking medications affecting brain activity (e.g., anti-epileptics, CNS-active drugs). Additional exclusions included alcohol or substance abuse, serious medical conditions (e.g., uncontrolled hypertension, cardiac arrhythmias, severe respiratory disorders), recent head injuries, untreated sleep disorders, pregnancy, severe psychiatric disorders (e.g., schizophrenia, bipolar disorder), and metal implants or devices that could interfere with EEG. Individuals who had previously watched any episodes of BBC’s Sherlock were also excluded to avoid bias. During data collection, one older adult was excluded for falling asleep, two young adults for excessive sweat artifacts, and one young adult for electrode malfunctions. The final sample comprised 84 participants: 42 young adults and 42 older adults.

Stimuli and tasks

Stimuli

Participants watched an audiovisual video clip while EEG data were recorded. To capture attention and ease the transition into the main stimulus, they first viewed a 1 min 2D animation (‘Morning!’) chosen for its distinct visual and narrative style, providing a baseline for engagement. This was followed by the first 35 min of BBC’s Sherlock (from the episode A Study in Pink), a stimulus previously used in naturalistic narrative studies^27,33,48,50.

Free recall task

In the free recall task, participants typed a description of the Sherlock episode in PsychoPy (v2023.2.3) without any cues. They were instructed to report events in order, include as many details as possible, and imagine explaining the episode to someone who would later be tested. They had 15 min to complete the task.

To define the events referenced in the free recall task, we generated a set of event boundaries using an independent online sample (n = 50). They identified 35 events (see Supplementary Materials for details). The events varied in length, ranging from 11 to 212 s (M = 55.70, SD = 46.45). See Fig. 6 for a visualization of the annotated event boundaries. Each event was assigned a descriptive label (e.g., ‘Press Conference’).

Fig. 6.

Visual representation of event segmentation based on human annotations. Each color-coded rectangle represents a distinct event, with the width of the rectangle corresponding to the event duration. The start and end points of each rectangle indicate the annotated event boundaries. Colors indicate event-duration ranges: 0–20 s, 20–40 s, 40–60 s, and > 60 s (as shown in the legend).

Each participant’s recall was scored for whether any part of each event was mentioned (1 = mentioned, 0 = not mentioned), regardless of the amount of detail provided. Event-level binary scores were used for event-level analyses, and the total number of recalled events (0–35) served as a continuous measure of overall free recall performance in subject-level analyses.

Other measures

Participants also completed a psychometric battery of tests, including those that assess processing speed tasks, episodic long-term memory, attention, vocabulary knowledge, and a conscientiousness questionnaire. They also completed other measures of memory for the Sherlock events, including cued recall, recognition and temporal order memory. These additional event-memory measures were not included in the main event-level analyses because they sampled only a subset of the 35 events and therefore did not provide complete event-level memory data for every event, as the free recall task did. As a result, they could not be used to directly test whether within-event neural stability predicted memory for each individual event. However, a detailed description of these tasks, along with descriptive statistics for young and older adults and correlations between task performance and EEG similarity measures, is provided in the Supplementary Materials (see Supplementary Tables S3 and S4).

Procedure

After signing a consent form, participants were fitted with a 64-channel EEG cap using a gel-based method for optimal conductivity. They were then taken to a sound-attenuating booth (WhisperRoom, Knoxville, TN) where they received on-screen study instructions in a softly lit environment. First, participants watched the one-minute video clip. On-screen reminders were given to minimize head, hand, and leg movements to reduce EEG artifacts. Participants then watched the 35 min clip of Sherlock while the EEG data was recorded. Afterward, participants took a 10 min break during which the EEG cap was removed and cleaned. They then proceeded to the behavioral testing station to complete the cognitive assessments. All tasks were administered in the following order: Free recall, cued recall, recognition memory, and temporal order memory (administered via PsychoPy, v2023.2.3). These were followed by the National Institutes of Health (NIH) Toolbox Cognition Battery⁵¹: Rey Auditory Verbal Learning Test (RAVLT) encoding, Pattern Comparison Processing Speed Test, Flanker Inhibitory Control, Picture Vocabulary Test, and RAVLT delayed test. Finally, participants completed a Qualtrics survey, were debriefed, and thanked. They were informed that standardized test scores and their conscientiousness score would be emailed after study completion.

EEG recording and preprocessing

EEG data were recorded using a BioSemi Active II system (BioSemi, Amsterdam, Netherlands) at 2048 Hz with 24-bit resolution. A 70-channel electrode array included a 64-electrode cap positioned according to the 10–20 system, with cap sizes chosen for each participant’s head. The remaining six electrodes were placed below each eye, at the temples, and on each mastoid. Signals were referenced online to the Common-Mode-Sense/Driven-Right-Leg (CMS/DRL) reference specific to the BioSemi system (additional details available at biosemi.com). To ensure data quality, DC offsets between CMS/DRL and each electrode were maintained below 25 μV; electrodes exceeding this threshold were excluded from analysis.

The video was presented on an AOC 24-inch monitor (model 12476VXM), and audio was played through Reftone LD-3 stereo speakers (Reftone, Woodland Hills, California) at ~ 70 dB SPL. After importing and preprocessing, the data were downsampled to 512 Hz, low-pass filtered at 50 Hz to reduce high-frequency artifacts^52–54 and high-pass filtered at 0.1 Hz to remove slow drifts caused by low-frequency noise (e.g., skin potentials, sweat artifacts) while preserving cognitive-relevant signals⁵⁵. Independent Component Analysis (ICA) in MATLAB’s EEGLAB toolbox⁵² corrected eye movements and removed noisy components (≤ 10 components per participant). The data were normalized prior to analysis by applying z-score normalization to all time points, ensuring that the mean activity of each electrode was zero across time. To evaluate whether preprocessing introduced age-dependent differences in data quality, we compared the number of ICA components removed and the number of interpolated scalp channels between age groups. The number of removed ICA components did not differ significantly between groups: young adults, M = 9.57, SD = 0.94, range = 6–10; older adults, M = 9.74, SD = 0.94, range = 5–10; Welch’s t(82) = − 0.81, p = 0.419. Older adults required significantly more interpolated scalp channels (M = 4.52, SD = 2.07, range = 0–8) than young adults (M = 2.98, SD = 1.97, range = 0–9), Welch’s t(81.78) = 3.51, p < 0.001, consistent with the greater EEG artifact burden often observed in aging.

Photosensor-based event boundary synchronization

At each event boundary, a masked flash appeared in the top-right corner of the screen, and a photosensor placed discreetly on the monitor recorded its timing. This photosensor data was added as a new channel to the EEG recordings, enabling precise tagging of event boundaries without distracting participants during movie viewing.

EEG pattern similarity within and across events

Trial-to-trial EEG segments were defined relative to event boundaries (Fig. 7). Data were epoched into two 3 s segments: one starting 3 s after each event boundary (“segment 1”) and the other ending 3 s before the next event boundary (“segment 2”). This fixed-segment approach was chosen to provide a standardized event-level estimate of neural stability while maintaining equal-length EEG vectors across participants and events. The two segments were positioned 3 s away from event boundaries to reduce contamination from immediate boundary-related neural responses, such as event model updating or boundary-triggered reinstatement. The segments were also selected to be non-overlapping, which was important because some events were relatively short. Thus, this approach was intended to capture broad similarity between two temporally separated moments within the same event, rather than fine-grained temporal dynamics across the entire event. Each segment was transformed into equal-length EEG vectors of 3 s (1536 time points at 512 Hz). Each 3-s EEG segment was recorded from all 64 electrodes and, at a sampling rate of 512 Hz, comprised 1,536 time points, produced a 64 × 1,536 matrix of voltage values. Similarity between paired segments was quantified by computing a Pearson correlation across the 64 electrodes at each time point, comparing the scalp-wide spatial voltage pattern in one segment with the corresponding pattern in the paired segment. The resulting 1,536 correlation coefficients were then averaged to obtain a single event-level similarity value. Representational Similarity Analysis was conducted at the individual level using spatial (all 64 electrodes) and temporal features. Similarity was computed via Pearson correlation, which is insensitive to absolute amplitude and variance.

Fig. 7.

Schematic of EEG epoching and similarity analysis. EEG data were segmented into 3-s epochs relative to event boundaries (EB), producing two segments per event (Segment 1 and Segment 2). Within-event similarity compares segments within the same event, while across-event similarity compares segments from adjacent events. Highlighted regions indicate the analyzed epochs.

Within-event correlations paired segment 1 with segment 2 from the same event (mean temporal distance = 43.74 s, range 4.6–205.76 s). Across-event correlations paired segment 2 of event n with segment 1 of event n + 1, separated by 6 s. Point-to-point correlations produced 2D trial-wise matrices for within- and across-event correlations. Mean and standard deviation of correlations were calculated per subject, and event-specific similarity values were used to compare neural similarity for later remembered versus forgotten events.

To control for low-level perceptual differences, perceptual change between segment pairs was quantified by averaging RGB values over 3 s segments (90 frames) and computing the normalized 3D Euclidean distance between segments^18,56,57. These values were included as covariates in statistical models to account for trial-by-trial perceptual variability.

To assess whether similarity estimates were reliable within each age group, we computed split-half reliability by correlating participant-level similarity estimates derived from odd- and even-numbered events. Within-event similarity showed strong reliability in both older adults (r = 0.733, Spearman–Brown corrected r = 0.846) and young adults (r = 0.733, Spearman–Brown corrected r = 0.846). Across-event similarity also showed acceptable reliability in older adults (r = 0.590, Spearman–Brown corrected r = 0.742) and young adults (r = 0.642, Spearman–Brown corrected r = 0.782). These results indicate that similarity estimates were reasonably stable in both age groups and do not suggest markedly lower measurement reliability in older adults.

Experimental design and statistical analysis

The current study used a 2 (Age Group: Young vs. Older) × 2 (Trial Type: Within-event vs. Across-event) mixed design, with Age Group as a between-subjects variable and Trial Type as a within-subjects variable, predicting either EEG neural stability (continuous variable) or free recall (both binary and continuous variables). Mixed-effects models were used to account for repeated measures nested within participants and events, with linear models for continuous outcomes and logistic models for binary outcomes. Models included random intercepts for participants and events. For analyses involving the within-subject factor Trial Type, we additionally evaluated by-participant random slopes for Trial Type. When random-slope models produced singular fits or did not change the substantive fixed-effect conclusions, the more parsimonious random-intercept models were retained for interpretation. Final models were selected using Akaike Information Criterion (AIC) and likelihood ratio tests; model fit details are provided with each analysis below.

To examine whether relative neural differentiation across event boundaries predicted memory for adjacent events, we also conducted a boundary-level binomial mixed-effects analysis in which recall for the events preceding and following each boundary was summed and predicted from a within-minus-across similarity differentiation score, Age Group, and their interaction. To further assess age-related differences in the magnitude of within-event neural stability, we computed a participant-level difference score by subtracting across-event similarity from within-event similarity for each participant. These difference scores were compared between age groups using Welch’s t-test, followed by equivalence testing and a Bayesian independent-samples t-test. These follow-up analyses were conducted to evaluate whether the absence of a significant Age Group × Trial Type interaction was consistent with equivalence or instead reflected inconclusive evidence regarding age-group differences.

To complement the frequentist analyses, we also computed Bayes factors for the major effects of interest. Bayes factors quantify the relative evidence for one hypothesis or model compared with another and are especially useful for interpreting non-significant frequentist results. We report BF₁₀ when evaluating evidence for the alternative hypothesis relative to the null hypothesis and BF₀₁ when evaluating evidence for the null hypothesis relative to the alternative hypothesis. Values greater than 1 indicate more evidence for the numerator hypothesis. Following common interpretive conventions, values between 1 and 3 were interpreted as anecdotal evidence, values between 3 and 10 as moderate evidence, and values greater than 10 as strong evidence.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

H.B. and M.D. designed the study. M.D. collected the data, performed the analyses, and drafted the manuscript. M.T. contributed analytic tools and assisted with data analysis. All authors contributed to interpretation of the results and approved the final version of the manuscript.

Funding

This work was supported by the National Institutes of Health (Grant R01AG075012-01A1 to H.B.) and by the National Institutes of Health, through the Cognitive and Neurobiological Approaches to Plasticity (CNAP) Center of Biomedical Research Excellence (Grant P20GM113109 to H.B.).

Data availability

The datasets generated and/or analyzed during the current study are available in the Open Science Framework (OSF) repository https://osf.io/k4cus/. The shared files contain the minimal de-identified dataset necessary to interpret, replicate, and build upon the findings reported in this article.

Declarations

Competing interests

The authors declare no competing interests.