Retention and transformation of internal experiences in autobiographical memory narratives

Abstract

Real-world memories often include our internal thoughts and feelings, yet memory research has largely centered on experimentally controlled external stimuli. As a result, it remains underexplored how such internal experiences are preserved or transformed in naturalistic autobiographical recall, and whether they serve functional roles. To address this gap, we analyzed a large-scale dataset of autobiographical narratives in which 210 crowdsourced participants recalled a specific memorable life event twice, several weeks apart, and rated their memories along multiple dimensions. We combined manual annotation with natural language processing to identify and analyze individual memory details, categorized as either observable external experiences or subjective internal experiences. We found that internal experiences were more prone to omission and semantic distortion over time compared to external ones. However, those with higher emotional intensity and stronger semantic connections to external event features were more likely to be retained. Memories richer in internal experiences were also judged as more important, and semantically precise reinstatement of internal experiences across recalls predicted increased importance over time. Together, these findings show that although memories of internal thoughts and feelings are relatively fragile, they nonetheless play a meaningful role in shaping the subjective significance of autobiographical memory.

Using repeated autobiographical recall, this study shows that internal thoughts and feelings are less stable over time than externally oriented details, yet their retention and stability contribute to the perceived importance of personal memories.

Introduction

Recalling internal experiences—our thoughts and feelings known only to ourselves—is a fundamental part of everyday memory. We reflect on our impressions of a movie when recommending it to a friend, revisit project ideas at work that came to us during our morning shower, or recount emotions from a past relationship in a therapy session. These memories of internal experiences, not just those involving externally observable facts, play a crucial role in guiding goal-directed behavior and decision-making^1,2, as well as in maintaining a coherent sense of self³. Self-reports of past thoughts and feelings are also widely used in clinical settings to assess mental health⁴. Yet, despite their ubiquity and functional significance, memories of internal experiences have received far less attention in memory research than memories of externally experienced stimuli and events⁵.

Previous research on memories of internal experiences has primarily examined experimentally induced thoughts and feelings. In these studies, participants are typically instructed to imagine specific stimuli or events^6–10, often designed to be comparable in content to actual experiences^11–13, and are then tested on their memories. This approach has shown that memories of imagined events contain less spatiotemporal and sensory information^11,12 and decay faster than memories of real events¹². In addition, imagined events are better remembered when they are plausible^13,14, detailed¹⁰, emotional¹⁵, self-relevant¹⁶, or involve familiar elements^10,15. However, the memory phenomena observed in these controlled laboratory experiments, constrained by specific task demands, may differ from those associated with real-world memories^17–19.

To investigate memories of thoughts and feelings that arise naturally in real life, another line of research has focused on autobiographical memory. These studies directly prompt participants to report memories of internal experiences that occurred during routine or memorable everyday events^20–23, major historical or political moments^24–27, or staged real-world scenarios such as walking around a university campus²⁸. Because these uncontrolled real-life experiences—especially internal ones—are difficult to objectively verify, researchers often assess the consistency of memories across repeated sessions rather than their accuracy^29,30. The initial report, made either immediately or within a relatively short period after the experience, is treated as a closer approximation of the original memory and used as a benchmark for later comparisons. In these studies, internal experiences are generally remembered less consistently than external aspects of events, such as actions^21,22,24. Memories of feelings, in particular, are susceptible to bias from one’s current emotions and evaluations of the past event^31–33.

However, relatively little is known about how memories of naturally occurring internal experiences are expressed and preserved over time within autobiographical narratives. This gap is notable because narrative—accounts of temporally unfolding, interrelated events involving agents, goals, and actions—serves as a primary format through which real-life episodes are stored, retrieved, and studied^34–38. In autobiographical memory research, participants typically generate narratives by describing, in their own words, what they remember about a specific life event^39–42. Unlike directly probing memories of internal experiences through isolated questions, the narrative method provides a more ecologically valid approach to examining thoughts and feelings embedded within broader contexts and their relationships to other memory details. It also helps reduce potential demand characteristics associated with direct questioning, which may be especially relevant for memories of internal experiences, given their previously reported vulnerability to distortion³³.

The present study aimed to characterize the retention and transformation of real-life internal experiences across repeated narrations of autobiographical events. Specifically, we addressed three questions: (1) Are memories of internal experiences more likely to be forgotten or altered over time than memories of external experiences in autobiographical narratives? (2) What features of internal experiences predict whether they are later remembered? (3) How does recalling internal experiences relate to people’s subjective perception or evaluation of past events? To investigate these questions, we leveraged a large-scale dataset of autobiographical narratives (the Hippocorpus dataset^43–45), in which crowdsourced participants described the same recent memorable event twice, weeks apart, and rated various aspects of their memories. Using a combination of manual annotation and automated natural language processing techniques^46,47, we identified and analyzed individual memory details embedded in the narratives, including internal thoughts and feelings as well as observable external elements. We hypothesized that memories of internal experiences would be more susceptible to forgetting and semantic distortion than memories of external experiences, since internal thoughts and feelings generally lack sensory or perceptual anchors¹¹ and are readily reshaped by current goals and appraisals, as shown in prior work^1,33. However, we predicted that emotional salience^48,49, informational richness^50,51, and semantic connections to external elements that may serve as concrete retrieval cues^52,53 would mitigate this vulnerability. We also hypothesized that more consistent and precise recall of private thoughts and feelings would be linked to greater perceived importance of the event, as such recollections are often emotionally charged and enhance the sense of reliving during retrieval⁵⁴—factors that predict perceived self-relevance and importance^55,56.

Methods

To examine how memories of naturally occurring internal feelings and thoughts are preserved or transformed over time through repeated narrative recall, we analyzed written autobiographical memory narratives from the publicly available ‘Hippocorpus’ dataset^43–45. We selected a subset of the data and conducted a series of text preprocessing steps, including segmenting narratives into individual memory details, categorizing each detail, and manually matching details across the two recall sessions. The primary analyses comparing memory details that described internal versus external experiences were preregistered on AsPredicted (https://aspredicted.org/9zmr-djp5.pdf; Preregistration date: March 2nd, 2024). Analyses examining (1) predictors of memory detail retention and (2) perceived memory importance were exploratory and not included in the original preregistration.

The Hippocorpus dataset

The Hippocorpus dataset^43,45 comprises 6854 first-person narratives describing recalled or imagined memorable life events, collected from crowdworkers on Amazon Mechanical Turk. The dataset was generated in accordance with protocols approved by the Institutional Review Board at Microsoft Research, and all participants provided written informed consent. Participants received monetary compensation for their time. Data collection involved two recall sessions spaced several weeks apart (Fig. 1). In the first session, participants were instructed to write a 15–25 sentence diary-like narrative about a memorable event that had occurred recently (approximately within the past three to 6 months). They were also instructed to provide a brief summary (2–3 sentences) of the event, including key details to help them remember what they had written when recalling the event in the future (see Supplementary Methods for analyses of the memory details contained in these summaries). In the second session, participants received their own summary as a cue and were asked to recall and describe the same event again. Additionally, a separate group of participants was given these summaries and asked to write imagined narratives based on them. These responses were not relevant to the current study and were excluded from analysis. Immediately following each recall task, participants completed a demographic questionnaire and provided responses to several questions regarding the recalled event and their mental state during the task. These included, but were not limited to, how long ago the event occurred, how important/impactful/personal it was, and how distracted they felt while writing. Importance and distraction were rated on a 5-point Likert scale (1 = not at all, 5 = very much). A complete list of the post-recall questions is available in ref. ⁴³.

Fig. 1. Schematic of the task and data preprocessing.

We analyzed written autobiographical memory recall data from the Hippocorpus dataset^43,45. In the first session, participants described a recent memorable event and provided a brief summary. In the second session, they recalled the same event using their summary as a cue. Responses were manually segmented into individual memory details and categorized as externally oriented, internally oriented, inferences about others, general comments, or other. In the illustrative example (not based on actual data), externally oriented, internally oriented, and other details are shown in green, orange, and gray, respectively. Details from both recalls were manually matched to determine whether each was retained, omitted, or added in the second recall.

Data selection

Since our study focused on the retention and transformation of memory across repeated recalls, we initially selected data from 1308 participants who described the same event in two recall sessions, with the summary generated in the first session provided as the cue for the second recall. Participants who provided imagined narratives or who completed only the first, but not the second, recall session were excluded from analysis.

We then applied additional step-by-step screening procedures based on criteria pre-defined in our preregistration protocol (https://aspredicted.org/9zmr-djp5.pdf), resulting in a final sample of 210 participants included in our analyses. First, to ensure recall quality, we selected 822 participants who were not distracted during memory recall. These participants responded “Not at all” to the question, “How distracted were you while writing your story?” during both the first and second recall sessions. Next, to focus on memories of clearly identifiable, specific events, we included only participants whose first and second recalls described the same specific event that occurred within a single day. Participants were excluded if they recalled different events across sessions, described situations that unfolded over several days or weeks (e.g., changes in life after returning to college), provided only general statements about recurring events, or produced narratives that were disorganized and lacked a clear episodic focus. This selection was performed manually by one of the authors (C.K.), who reviewed the content and quality of both recalls. This step resulted in a remaining pool of 330 participants. Finally, we selected 210 participants who recalled an event that occurred within three months (90 days) prior to the first recall session. This criterion was intended to focus on relatively recent memories from the Hippocorpus dataset^43,45, as older, more stabilized memories typically allow less room for transformation between recall sessions^30,57.

Based on data collected during the first recall session, the final sample of 210 participants included 14 aged 18–24, 42 aged 25–29, 49 aged 30–34, 38 aged 35–39, 15 aged 40–44, 19 aged 45–49, 17 aged 50–54, and 16 aged 55 or older. Of these participants, 118 identified as female, 90 as male, 1 as nonbinary, and 1 did not report their gender. Participants also self-reported their race/ethnicity: 156 identified as White, 17 as Black or African American, 14 as Hispanic/Latinx, 11 as Asian or Asian American, 7 as Mixed/Other, 2 as Indian American/South Asian, and 1 participant each identified as Middle Eastern, Native American/Hawaiian, or “prefer not to disclose.” On average, recalled events were 41.8 days old (SD = 29.3, range: 7–90) at the first session and 126.6 days old (SD = 71.1, range: 21–540) at the second session.

Memory detail segmentation

The selected narratives from both recall sessions underwent a series of preprocessing steps in preparation for analysis. First, each narrative was segmented into individual memory details, which served as the primary units of analysis (Fig. 1). This segmentation was performed manually by eight human annotators. Each narrative was segmented by a single annotator, with each annotator processing an average of 52.5 narratives (SD = 45.1, range: 4–149). Participants were randomly assigned to annotators. For all but one participant, the same annotator segmented both the first and second recall sessions. Annotators were provided with text files containing the recall narratives, formatted so that each sentence appeared on a separate line. They were instructed to further divide these sentences into smaller units, ensuring that each line reflected a single idea, without modifying the original content or wording. Specifically, sentences containing multiple distinct ideas were split into separate clauses. For example, the sentence “It was really nice outside, which made me happy.” would be divided into two lines: “It was really nice outside” and “which made me happy.” Clauses have commonly served as the standard unit of analysis in the literature on narrative comprehension and memory^58,59. Thus, to preserve the integrated, cohesive idea conveyed within each clause, we did not further subdivide them into finer-grained pieces of information.

Memory detail categorization

After segmentation, each memory detail from the recall narratives was manually categorized into specific types by five of the human annotators who had performed the segmentation. Each narrative was categorized by a single annotator, with each annotator processing an average of 84 narratives (SD = 88.5, range: 2–200). For consistency, the same annotator categorized both the first and second recall sessions for each participant. Participants were randomly assigned to annotators. Annotators received spreadsheets containing the segmented recall narratives, with each memory detail listed in a separate row. In an adjacent column, they assigned a category to each detail by entering a number corresponding to one of the predefined categories (see Supplementary Table 1 for an example). Annotators used their best subjective judgment, classifying each detail according to its dominant meaning or focus. The categories were as follows: (1) specific objective external experiences (externally oriented), (2) specific subjective internal thoughts or feelings (internally oriented), (3) speculations or inferences about others, and (4) general comments or evaluations about the recalled event. Any information not directly related to the main recalled event was categorized as ‘other details.’

Annotators were also provided with detailed definitions and examples for each category. The first category, specific objective external experiences (externally oriented details), included observable sensations, actions, and other concrete details that the participant directly experienced during the unfolding of the recalled event (e.g., “We stumbled upon an abandoned house while walking in the woods”). Only externally observable and event-specific experiences were included, while general facts or knowledge about the participant, other people, or the world were excluded.

The second category, specific subjective internal thoughts or feelings (internally oriented details), encompassed the participant’s personal emotions, thoughts, intentions, or interpretations experienced during the recalled event. These were internal states accessible only to the participant (e.g., “I was very nervous and I thought I wouldn’t go inside the house”). This category was limited to internal experiences that occurred during the event itself and excluded broader personal beliefs or general tendencies (e.g., “I always feel happy when I am holding a puppy”). Similarly, post-event reflections such as “It was a memorable event” were not included in this category and were instead classified as general comments or evaluations (see below).

The third category, speculations and inferences about others, included the participant’s interpretations or assumptions about other people’s thoughts, feelings, or intentions during the recalled event (e.g., “My son wasn’t aware of what was about to happen”), in contrast to their own internal states. Descriptions in this category were typically based on inferences drawn from others’ observable behaviors or expressions, as participants cannot directly access another person’s internal experiences. However, direct reports of others’ statements or actions (e.g., “She said she was happy”) were not included in this category; instead, they were classified as specific objective external experiences.

The fourth category, general comments or evaluations, encompassed reflections, interpretations, or assessments made about the event during the memory recall sessions (e.g., “It was one of the happiest moments of my life”). A key distinction between this category and subjective internal thoughts or feelings lies in the timing of the reflection. For example, the statement “I had a great time” would be classified as a general comment or evaluation, as it reflects a post-event assessment made after the experience had ended. In contrast, “I was having a great time” would be categorized as a subjective internal thought or feeling, as it describes an emotional experience occurring during the event itself.

Finally, details that did not align with any of the categories described above were classified as ‘other.’ These were not directly related to the main recalled event and included, but were not limited to, general semantic knowledge about oneself or others (e.g., “He is normally a happy guy”), meta-memory comments (e.g., “I can’t remember”), thoughts or feelings about the future related to the recalled event (e.g., “I will have to get another job”), and recollections of past events other than the main event being recalled (e.g., “Last year I didn’t have a very good birthday”).

Since the primary goal of the current study was to examine the retention and transformation of internal thoughts and feelings over repeated recalls and to compare these patterns with those of objective external elements, these two categories were defined and analyzed as the primary categories of interest. Given our focus on the broad internal-external distinction, we did not further subdivide externally oriented details into finer-grained categories (e.g., time, location, and perception), as is done in previously established protocols (e.g., ref. ³⁹). The categories of ‘speculations or inferences about others’ and ‘general comments or evaluations’ were additionally introduced to help annotators avoid misclassification. The former distinguished other people’s inferred internal experiences from participants’ own, while the latter separated thoughts or feelings generated during the act of recall from those that occurred during the original event. Because internally oriented details and these additional types of details all described internal states using similar language, they were easily confused. By explicitly flagging these categories for annotators and instructing them to classify such details separately, we aimed to more effectively identify and exclude them from the main analysis—together with the ‘other’ category—as they were not relevant to the study’s primary goal.

It is also important to note that our internally and externally oriented memory detail categories differ from the “internal” and “external” details defined in the widely used Autobiographical Interview (AI) protocol³⁹. In the AI protocol, internal details refer to episodic elements directly related to the main recalled event, whereas external details encompass information not directly tied to that event (e.g., semantic knowledge). Under this framework, both the internally oriented and externally oriented details in our study correspond to the AI protocol’s internal details, while our ‘other’ category corresponds to the AI protocol’s external details.

Matching details across recall sessions

The final preprocessing step involved matching segmented memory details from the first and second recalls (Fig. 1) to determine which details were retained, omitted, or newly added in the second recall. This matching was manually performed by four of the human annotators who had also segmented and categorized the memory details. Each pair of narratives was matched by a single annotator, with each annotator processing an average of 52.5 narrative pairs (SD = 32.7, range: 18–90). To promote efficiency, annotators were assigned to match narratives from the same participants they had categorized whenever possible, resulting in 179 of 210 participants’ narratives being categorized and matched by the same annotators.

Annotators were given two spreadsheets per participant, one for each recall session, with each memory detail listed in a separate row. Details from the first recall were numbered sequentially. For each detail in the second recall, annotators recorded the number(s) of the corresponding detail(s) from the first recall, if a match was identified (see Supplementary Table 2 for an example). If no match was found, the cell was left blank. Thus, the matching outcome was binary: a detail from the first recall either matched or did not match a detail from the second recall; no intermediate category, such as “similar but not identical,” was allowed. Annotators were instructed to use their subjective judgment to pair only those details that described the same content (e.g., actions and people). Exact wording was not required, but the meaning had to be equivalent or highly similar. For example, “I started crying and said, ‘yes, definitely yes!’” and “Of course I said yes!” were considered a match. Cross-session semantic similarity was higher for matched compared to non-matched details (Supplementary Fig. 1), confirming that annotators were indeed pairing details that were similar in meaning.

Once the matching procedure was complete, details from the first recall that had corresponding matches in the second recall were classified as retained, whereas those without a match were classified as omitted. Likewise, details in the second recall that matched details from the first recall were categorized as retained, and those without a match were classified as newly added. The retention rate was calculated as the number of first-recall details with a corresponding match in the second recall divided by the total number of details produced in the first recall. Retention rates were also computed separately within each detail category. For memory addition, we calculated the proportion of each detail category within the newly added details. Specifically, for each category, this proportion was defined as the number of newly added details of that category in the second recall divided by the total number of added details.

For all steps of recall narrative preprocessing, annotators were trained using a shared set of written instructions and were provided with brief hypothetical examples (see Supplementary Table 1 for a memory categorization example) to ensure consistency across annotators. Their performance was monitored by one of the authors (C.K.), and they were encouraged to ask questions before and during the task to clarify any uncertainties. C.K. also conducted a quality check of each completed recall narrative and requested revisions when any issues were identified. In addition, we retrospectively assessed the inter-rater reliability of memory detail categorization by having two additional independent annotators code a randomly selected subset of data from 20 participants. The overall Fleiss’s Kappa across the three annotators (including the original annotator) was 0.54, indicating moderate agreement. Category-specific Kappa values were 0.64 for externally oriented, 0.57 for internally oriented, 0.54 for inferences about others, 0.39 for general comments, and 0.42 for other categories.

Measuring semantic similarity across recall sessions

To assess how precisely a memory detail was preserved over time, we measured the semantic similarity between matched memory details from the first and second recalls using a natural language processing approach⁴⁶. Each memory detail was converted into a 768-dimensional embedding vector using a pretrained SBERT model⁶⁰ (all-mpnet-base-v2) implemented in the Sentence Transformers Python package (version 2.2.0, https://www.sbert.net). The model is trained so that texts with similar semantic content produce embedding vectors that align closely in the multidimensional space, regardless of whether the information conveyed is abstract or concrete. Cross-session semantic similarity was then calculated as the cosine similarity between the embedding vectors of each matched pair. When a single memory detail from the first recall corresponded to multiple details in the second recall, we computed similarity using the concatenated text of all matched second-recall details. These semantic similarity values were then averaged across all matched pairs within each memory detail category to generate a summary score for each category and participant for comparison across categories.

Additionally, we examined cross-session semantic similarity values after accounting for the number of words in each memory detail pair. Prior research suggests that longer recall descriptions tend to be associated with higher semantic similarity⁴⁰. To control for this effect, we regressed out the total word count of each memory detail pair (i.e., the sum of words in the matched first and second recall details) from their semantic similarity scores, separately within each participant and across all detail categories. The resulting residuals—cross-session semantic similarity values with word count effects removed—were then averaged within each memory detail category to produce summary scores.

Predicting the retention of individual memory details

We conducted mixed-effects logistic regression analyses to examine which features of individual memory details from the first recall predicted their likelihood of being retained in the second recall. Since the Hippocorpus dataset contained no information about individual memory details beyond the recalled text itself, all predictors had to be features that could be extracted through text analysis. Among these text-derived features, we focused on those likely to influence memory retention based on prior findings^49,50,52, including emotional valence, emotional intensity, semantic connection to external elements, and word count per detail.

We selected emotional valence and intensity as predictors because prior work has shown that emotional autobiographical memories are generally remembered better than neutral ones^40,49,61. The emotional valence and emotional intensity of each memory detail were automatically rated using a natural language processing approach. Specifically, we applied the rule-based VADER (Valence Aware Dictionary and sEntiment Reasoner) sentiment analysis tool⁶², implemented in the Natural Language Toolkit (NLTK) Python library (version 3.8.1; https://www.nltk.org). VADER relies on a large lexicon of words that were empirically rated by independent human judges, with each word assigned a valence score. These scores are then adjusted using a set of linguistic rules, such as handling negations (e.g., “not good”), intensifiers (“very good”), punctuation, and capitalization. VADER combines and normalizes this information to generate a single polarity score for each text (i.e., memory detail), ranging from −1 to 1, where negative values indicate more negative sentiment and positive values indicate more positive sentiment. Emotional valence was represented directly by these polarity scores. Emotional intensity was calculated as the absolute value of the polarity score, with higher values indicating stronger emotional expression, regardless of whether the sentiment was positive or negative. Notably, these scores provide surface-level estimates of expressed affect for individual memory details and do not incorporate broader contextual interpretation.

Additionally, we examined the semantic connections between each memory detail and other externally oriented details. Earlier research has demonstrated that rich semantic connections among external events can enhance recall for those events⁵². For each internally oriented or externally oriented memory detail recalled during the first session, we computed its semantic similarity to external elements from the same session using the SBERT model⁶⁰ described above. To examine the effects of different levels of external context, we calculated both local and global semantic similarity. Local similarity was defined as the average cosine similarity between the embedding vector of the target detail and those of its immediately preceding and following externally oriented details. If the adjacent details were not externally oriented, the nearest preceding and following externally oriented details were used instead. Global similarity was defined as the average cosine similarity between the target detail and all other externally oriented details from the session.

Finally, we included word count per detail as a proxy for how elaborate or informationally rich each memory detail was. More elaborated memories with richer features are generally remembered better^50,51. For each detail, we counted the total number of words and did not exclude function words such as articles and prepositions. This approach preserved the natural form of participants’ written responses and maintained consistency with other text-based analyses in the current study. Notably, the mixed-effects logistic regression results were qualitatively unchanged when we instead excluded function words and used the number of content words as the predictor (Supplementary Methods).

All mixed-effects logistic regression models were conducted separately for internally oriented and externally oriented details, with a binary dependent variable indicating whether each memory detail was recalled again in the second session. We first evaluated each of the five predictors—emotional intensity, emotional valence, global similarity, local similarity, and word count—in separate models that included the predictor as a fixed effect and participant as a random effect. To assess the unique contributions of the features, we also fit models including three conceptually independent predictors (emotional intensity, local similarity, and word count) as fixed effects, with participant again as a random effect. All fixed-effect predictors were z-scored across all relevant memory details (either internally oriented or externally oriented), pooled across participants. Statistical significance of each predictor was determined using a likelihood ratio test, comparing the full model (with all fixed and random effects) to a null model that excluded the predictor of interest. When individual predictors were tested in separate models, we applied a Benjamini–Hochberg false discovery rate (FDR) correction across the five likelihood ratio tests to adjust the resulting p values for multiple comparisons.

Statistical tests

We used paired-samples t-tests to compare means between two conditions. For comparisons involving three or more conditions, repeated-measures ANOVAs were conducted. Mauchly’s test was used to assess the assumption of sphericity, and Greenhouse-Geisser corrections were applied when this assumption was violated. Pearson correlations were used to examine linear relationships between variables, with each participant contributing one data point. Differences between correlation coefficients were tested for statistical significance using the Williams test. All statistical tests were two-tailed. For any within-participant comparison between conditions, participants with missing data in any relevant condition (e.g., participants with no internally oriented details when comparing cross-session semantic similarity between internally oriented and externally oriented details) were excluded from the analysis.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Proportion of different memory detail categories

We first examined the presence and proportion of internal thoughts and feelings in autobiographical narratives written by 210 participants from the Hippocorpus dataset^43,45. Participants recalled a memorable life event twice, separated by an average of 84.8 days (SD = 61.9; see Supplementary Fig. 2 for the distribution of delay durations). Each recalled narrative was manually segmented into individual memory details, with each detail representing a single idea (Fig. 1). In the first recall, participants produced an average of 26.8 memory details (SD = 8.1) comprising 271.9 words (SD = 86.7). In the second recall, they produced an average of 25.0 details (SD = 6.4) comprising 252.3 words (SD = 75.9). Both the number of memory details and total word count were greater in the first recall (memory detail t(209) = 3.61, p < 0.001, 95% CI = [0.79, 2.70], Cohen’s d = 0.24; word count t(209) = 3.72, p < 0.001, 95% CI = [9.20, 29.97], Cohen’s d = 0.24).

Independent annotators manually assigned each memory detail to one of the following categories: externally oriented (e.g., “There was a DJ playing music”), internally oriented (e.g., “I thought I might faint”), speculations or inferences about others (e.g., “He seemed to be happy”), general comments or evaluations about the event (e.g., “It was a memorable day”), and other details not directly related to the recalled event. Externally oriented details made up the largest proportion of memory details in both the first recall (M = 57.1%, SD = 19.8) and the second recall (M = 59.4%, SD = 19.5), followed by internally oriented details (first: M = 10.9%, SD = 9.1; second: M = 10.5%, SD = 9.4), general comments (first: M = 7.2%, SD = 6.7; second: M = 7.0%, SD = 6.8), and inferences about others (first: M = 1.7%, SD = 3.1; second: M = 2.1%, SD = 3.9). Details unrelated to the recalled event accounted for 23.0% (SD = 18.0) and 21.0% (SD = 18.6) of the first and second recalls, respectively (Fig. 2). These differences in proportion across categories were statistically significant (F(4, 836) = 665.04, Greenhouse–Geisser corrected p < 0.001, η_p² = 0.76, 90% CI = [0.73, 0.78]), with no significant interaction between category and recall session (F(4, 836) = 2.28, Greenhouse–Geisser corrected p = 0.11, η_p² = 0.01, 90% CI = [0.00, 0.02]).

Fig. 2. Percentage of memory details by category.

Dark blue and magenta circles indicate the mean percentage of each detail category produced during the first and second recall sessions, respectively. Light blue and pink dots represent individual participants (N = 210 per category per session). Error bars denote 95% confidence intervals across participants.

Further analysis focusing on the externally oriented and internally oriented categories revealed that all 210 participants recalled at least one externally oriented detail in both sessions. In contrast, 179 participants recalled at least one internally oriented detail in the first recall session, and 168 did so in the second session. Similarly, the proportion of externally oriented details was higher than that of internally oriented details in both the first recall (t(209) = 27.56, p < 0.001, 95% CI = [42.89, 49.50], Cohen’s d = 3.00) and the second recall (t(209) = 29.86, p < 0.001, 95% CI = [45.66, 52.11], Cohen’s d = 3.19). These results indicate that while most participants did recall internal experiences, such details were less frequent than those describing external experiences. Given our focus on comparing memories of internal experiences with those of external experiences, the following analyses concentrate on these two categories. Results for the remaining categories are reported in the Supplementary Information (Supplementary Tables 3, 4, and 6).

Retention and addition of memory details

We next examined how memory details describing external and internal experiences changed across repeated recalls. To do this, we manually matched memory details from the first and second recall sessions based on their content (Fig. 1). First, as a measure of memory retention or omission, we calculated the proportion of memory details reported during the first recall that were also recalled during the second session. On average, 41.3% (SD = 15.3) of all detail types reported in the first recall were retained in the second. Importantly, when externally and internally oriented details were analyzed separately (Fig. 3a), internally oriented details (M = 38.3%, SD = 32.7) were significantly less likely to be retained than externally oriented details (M = 48.9%, SD = 19.2), indicating that internal experiences are more susceptible to omission over time (t(178) = 3.73, p < 0.001, 95% CI = [4.98, 16.19], Cohen’s d = 0.39). This pattern remained qualitatively unchanged after excluding memory details mentioned in the brief summary sentences used as recall cues (see Supplementary Methods). The retention rate was also not significantly influenced by the length of the delay between the two recall sessions for either internally oriented or externally oriented details (see Supplementary Methods).

Fig. 3. Transformation of memories for external and internal experiences.

a Mean percentages of externally and internally oriented memory details from the first recall that were retained in the second recall (N = 210 and 179, respectively). b Mean percentages of externally and internally oriented details among newly added details in the second recall (N = 209 for both categories). c Mean semantic similarity between matched externally and internally oriented details across recall sessions (N = 209 and 127, respectively), computed as cosine similarity between SBERT-generated text embeddings. In all panels, black circles indicate group means, colored dots represent individual participants, and error bars show 95% confidence intervals. Statistical significance was assessed using two-tailed paired t-tests among participants with data for both categories. ***p < 0.001.

Although second recalls were generally shorter than the first—indicating greater omission—they also included new memory details not reported in the initial recall, consistent with findings from prior studies^30,63. To assess this aspect of memory transformation, we examined whether internal experiences were more or less likely to be added in the second recall compared to external experiences (Fig. 3b). On average, 55.7% (SD = 17.6) of the details reported in the second recall were newly introduced. Among these new details, internally oriented details comprised a significantly smaller proportion (M = 10.5%, SD = 11.6) than externally oriented details (M = 54.0%, SD = 24.8; t(208) = 21.05, p < 0.001, 95% CI = [39.45, 47.60], Cohen’s d = 2.25). This pattern mirrors the overall proportions reported in Fig. 2, indicating that the relative scarcity of internally oriented details is evident not only in the details recalled initially but also in the new details introduced during the second recall. However, within each category, the proportion of newly added details relative to all details of the same type did not differ significantly between internally oriented (M = 43.9%, SD = 39.8) and externally oriented details (M = 48.6%, SD = 20.6; t(209) = 1.68, p = 0.094, 95% CI = [−0.01, 0.10], Cohen’s d = 0.12).

Semantic similarity across recall sessions

Thus far, we have analyzed the binary presence or absence of specific memory details. However, even when a detail was present in both recalls, its precise meaning may have shifted over time, reflecting semantic distortion. To capture these more subtle and graded changes in memory, we assessed the semantic similarity between the text descriptions of retained details from the first and second recall. Specifically, we converted each detail into a sentence embedding using the SBERT language model⁶⁰ and computed cosine similarity between corresponding embeddings. When analyzed separately for externally and internally oriented details (Fig. 3c), internally oriented details showed significantly lower semantic consistency across recalls (M = 0.54, SD = 0.15) compared to externally oriented details (M = 0.60, SD = 0.10; t(125) = 4.04, p < 0.001, 95% CI = [0.03, 0.09], Cohen’s d = 0.45). This difference remained significant even after controlling for the number of words included in the details (internally oriented residual M = −0.04, SD = 0.12; externally oriented residual M = 0.01, SD = 0.05; t(125) = 3.94, p < 0.001, 95% CI = [0.03, 0.08], Cohen’s d = 0.57). These results suggest that internal experiences were not only less likely to be retained but also more prone to semantic distortion when they were remembered. For cross-session semantic similarity, retention, and addition metrics for categories beyond externally and internally oriented details, see Supplementary Table 3.

Features predicting memory detail retention

If memories of internal experiences are more susceptible to omission and semantic distortion than memories of external experiences, what factors might make them more durable? We investigated whether specific features of individual memory details from the first recall predicted their likelihood of being retained in the second recall. First, we examined the emotional properties of each detail (Fig. 4a), given their known influence on autobiographical recall^48,49. Emotional valence and intensity were estimated using polarity scores generated by a text-based sentiment analysis tool⁶², where the sign of the score reflected valence, and the absolute value represented intensity. Second, we assessed local and global semantic similarity with external experiences (Fig. 4b), based on the hypothesis that such connections could serve as concrete retrieval cues facilitating recall^52,53. Local similarity—between a detail and its nearest externally oriented details—indexes its connection to the immediate context, whereas global similarity—between a detail and all externally oriented details from the same recall—reflects its integration into the broader theme or schema of the event. Finally, we used the word count of each detail as a proxy for informational richness or complexity, assuming that more elaborated details would be more memorable⁵⁰. Overall, compared to externally oriented details, internally oriented details were more emotionally intense, consistent with their definition, but showed weaker semantic links to externally oriented details and were shorter in length (Table 1). Feature values for other categories of memory details are reported in Supplementary Table 4.

Fig. 4. Predicting memory detail retention from first-recall features.

a Emotional valence and intensity were derived using a sentiment analyzer from the Natural Language Toolkit in Python. Each memory detail received a polarity score ranging from −1 to 1, with negative and positive values indicating negative and positive valence, respectively. The absolute value of the score reflected emotional intensity. b Semantic connection to external experiences was measured using cosine similarity of SBERT-based text embeddings. For each detail, local similarity was defined as the average similarity to its closest externally oriented details. Global similarity was defined as the average similarity to all externally oriented details in the recall. In (a, b), illustrative examples of externally and internally oriented details are shown in green and orange, respectively. c, d Parameter estimates from mixed-effects logistic regression models predicting the retention of internally oriented (c N = 600 memory details from 179 participants) and externally oriented (d N = 3209 memory details from 209 participants) details using different memory detail features. Circles represent parameter estimates, and error bars indicate 95% confidence intervals. Statistical significance for each feature was assessed with a likelihood ratio test, comparing the full model to the null model excluding the feature of interest. Supplementary Table 5 reports the full statistics for each feature. *p < 0.05, **p < 0.01, ***p < 0.001 (FDR corrected).

Table 1.

Comparison of selected features between externally and internally oriented details recalled during the first session

Feature	Externally oriented mean (SD)	Internally oriented mean (SD)	Externally oriented vs. internally oriented comparison
			t	df	p*	95% CI	Cohen’s d
Emotional valence	0.10 (0.15)	0.15 (0.35)	−1.79	178	0.075	[−0.09, 0.00]	0.17
Emotional intensity	0.21 (.10)	0.41 (.19)	−11.87	178	<0.001	[−0.22, −0.16]	1.24
Global semantic similarity	0.28 (0.06)	0.24 (0.06)	8.60	177	<0.001	[0.03, 0.05]	0.69
Local semantic similarity	0.34 (0.06)	0.28 (0.08)	8.62	177	<0.001	[0.04, 0.07]	0.83
Word count	10.4 (2.3)	9.6 (3.9)	2.45	178	0.015	[0.14, 1.32]	0.23

^*Uncorrected two-tailed p value.

To test whether these features predicted the likelihood that an internally oriented detail would be retained or omitted during the second recall, we conducted mixed-effects logistic regression analyses. Because emotional valence and intensity were correlated (r(598) = 0.34, p < 0.001, 95% CI = [0.27, 0.41]), as were local and global semantic similarity with externally oriented details (r(598) = 0.66, p < 0.001, 95% CI = [0.61, 0.70]), we first tested each feature in a separate model, with the feature as a fixed effect and participant as a random effect (Fig. 4c). Likelihood ratio tests showed that local similarity, global similarity, and emotional intensity each positively predicted the retention of internally oriented details, whereas emotional valence and word count were not significant predictors (see Supplementary Table 5 for full statistics). We also selected three conceptually independent predictors—emotional intensity, semantic similarity (local), and word count—and examined their unique contributions by including them together in a single model. Both emotional intensity (β = 0.19, 95% CI = [0.02, 0.36], χ²(1) = 4.93, p = 0.026) and local similarity (β = 0.26, 95% CI = [0.09, 0.44], χ²(1) = 8.84, p = 0.003) remained significant predictors of internally oriented detail retention.

Are these effects of memory detail features specific to internal experiences, or do they generalize to external experiences? To address this question, we repeated the logistic regression analyses on externally oriented details. When tested individually in separate models, all features except emotional valence positively predicted the retention of externally oriented details (Fig. 4d; Supplementary Table 5). In the combined model with the three independent predictors listed above, both word count (β = 0.26, 95% CI = [0.18, 0.34], χ²(1) = 41.74, p < 0.001) and local similarity (β = 0.14, 95% CI = [0.06, 0.22], χ²(1) = 12.07, p < 0.001) remained significant, while emotional intensity did not (β = 0.05, 95% CI = [–0.03, 0.12], χ²(1) = 1.35, p = 0.24). Overall, semantic connections to external context supported the retention of both detail types, while emotional intensity more strongly influenced internally oriented details and informational richness more strongly influenced externally oriented details.

Perceived importance of memory

Our findings so far suggest that, although memories of internal experiences are more fragile than those of external ones, they tend to be more emotionally charged. Prior work shows that emotion enhances phenomenological qualities of memory, such as vividness, reliving, and perceived importance^55,56,64. This raises the question of whether remembering internal thoughts and feelings is associated with how people evaluate the significance of a recalled event. To test this, we used participants’ self-reported ratings of the perceived importance of their memories (i.e., how impactful, significant, or personal the event felt) on a 5-point scale, as provided in the Hippocorpus dataset^43,45.

First, we examined correlations between the relative proportion of internally and externally oriented details and participants’ perceived importance of the memory (Fig. 5a, b). A higher proportion of internally oriented details was associated with greater perceived importance, numerically in the first recall (r(208) = 0.11, p = 0.102, 95% CI = [−0.02, 0.24]) and significantly in the second recall (r(208) = 0.16, p = 0.024, 95% CI = [0.02, 0.29]). In contrast, a greater proportion of externally oriented details was significantly associated with lower importance ratings in both the first (r(208) = −0.17, p = 0.016, 95% CI = [−0.29, −0.03]) and second recalls (r(208) = −0.23, p < 0.001, 95% CI = [−0.35, −0.10]). Williams tests confirmed that the correlations for internally versus externally oriented details differed significantly in both the first (t(207) = 2.51, p = 0.013, 95% CI = [0.06, 0.49], Cohen’s q = 0.28) and second recalls (t(207) = 3.62, p < 0.001, 95% CI = [0.18, 0.59], Cohen’s q = 0.39). We also examined memory details categorized as “general comments,” which often included overall evaluations of the event, but their proportions were not significantly correlated with importance ratings (Supplementary Table 6).

Fig. 5. Relationship between perceived memory importance and the recall of external and internal experiences.

a, b Correlations between importance ratings and the proportion of externally and internally oriented details recalled during the first (a) and second (b) sessions (N = 210 per category and session). c Correlations between the change in importance ratings (second minus first recall) and the mean semantic similarity of matched details across sessions, analyzed separately for externally oriented (N = 209) and internally oriented (N = 127) details. In all panels, green and orange dots represent individual participants for externally and internally oriented details, respectively. Lines show best-fit regression lines, with shaded areas indicating 95% confidence intervals.

The results above suggest that participants reported more internal experiences in memories they perceived as more important. However, the perceived importance of a memory may change over time as memories fade or transform. Indeed, importance ratings were significantly higher in the first recall (M = 4.33, SD = 0.90) than in the second (M = 4.02, SD = 1.03; t(209) = 5.82, p < 0.001, 95% CI = [0.20, 0.41], Cohen’s d = 0.31). To explore this further, we examined whether changes in importance ratings across the two recalls were related to previously assessed metrics of memory transformation: the proportion of details retained, the proportion of added details, and the semantic similarity across sessions, analyzed separately for internally and externally oriented details. We found no significant relationship between changes in importance and either the proportion of retained or added details, for both internally and externally oriented details (Supplementary Table 7). However, greater semantic similarity of internally oriented details across sessions was associated with an increase in perceived importance (r(125) = 0.19, p = 0.037, 95% CI = [0.01, 0.35]), whereas no such relationship emerged for externally oriented details (r(207) = −0.004, p = 0.951, 95% CI = [−0.14, 0.13]; Fig. 5c). Although the difference between the correlations for internally and externally oriented details did not reach statistical significance (t(123) = 1.89, p = 0.061, 95% CI = [−0.01, 0.41], Cohen’s q = 0.21), the positive correlation for internally oriented details suggests that consistent and precise retention of internal experiences may help buffer against the general decline in perceived importance over time.

Discussion

The present study investigated how naturally occurring internal experiences—our private thoughts and feelings—are recalled over time within autobiographical narratives. Using natural language processing to analyze a large-scale dataset of autobiographical narratives (the Hippocorpus dataset^43,45), we found that internal experiences were more susceptible to omission and semantic distortion than observable external experiences. However, internal experiences were more likely to be retained when they were emotionally intense or semantically linked to externally oriented event details. Despite their fragile nature, the precise recall of internal experiences was also associated with greater perceived significance of autobiographical memories.

Internal thoughts and feelings have rarely been the central focus of cognitive psychology research on memory using autobiographical narratives, even though they are commonly recalled and explicitly included as a distinct category in standardized coding protocols such as the Autobiographical Interview³⁹. Prior studies using these protocols have typically reported the overall proportion and reliability of thought- and feeling-related details as one category among many^30,39,49. However, because internal experiences were not the primary target of this work, their features and characteristics at the level of individual details have received relatively little attention. The present study extends this prior research by conducting more fine-grained analyses of internal experiences within narratives, focusing on individual details and identifying the factors that contribute to their relative durability. Additionally, previous research assessing the accuracy of memories of real-world internal experiences—and comparing them with external experiences—has often relied on small samples, sometimes including the experimenters themselves^21,22,65, or has focused on highly selective emotional events such as the 9/11 attacks^24,32,57. By drawing on a large and diverse set of autobiographical narratives⁴³, the present study provides a more generalizable investigation of internal experiences in naturalistic memory, addressing their occurrence, stability, and relationship to the perceived importance of memory.

The accuracy of remembering internal thoughts and feelings has been debated. Some researchers argue that fleeting internal experiences do not persist in memory⁶⁶, whereas others suggest that they can be stored, exerting lasting effects and serving as efficient affective summaries of experience without preserving every detail^67,68. Our findings indicate that internal experiences are frequently remembered in everyday life, but they are also more vulnerable to omission and semantic distortion than external experiences, in line with prior empirical work^21,22,24. A recent study examining repeated autobiographical recalls similarly found that thoughts and emotions were recalled less consistently than most other types of memory details³⁰. Unlike externally oriented events, which benefit from multisensory cues (e.g., visual, auditory, and spatial) that provide multiple retrieval pathways, internal states generally lack such direct perceptual anchors^11,32. They are also less likely to form rich, image-based traces that support episodic recollection²². As a result, recalling internal experiences often involves inference: people reconstruct past thoughts or feelings based on semantic self-knowledge (e.g., “I always get nervous at parties”) or on current appraisals (e.g., “I dislike X, so I must have felt angry”), rather than retrieving the original experience⁶⁶. This reliance on reconstruction may make internal memories particularly susceptible to bias, as they are often reshaped by current beliefs and goals, as demonstrated in prior work^1,32,33.

Certain internal experiences were more likely to be retained over time, particularly those marked by high emotional intensity and strong semantic connections to externally oriented event content. The observed influence of emotional arousal aligns with its established role in supporting memory retrieval and consolidation^69–71. Prior studies have also shown that emotional autobiographical events are remembered with greater detail and vividness than neutral ones^40,49,61. Similarly, emotion has been found to enhance the consistency of autobiographical recall across repetitions, improving both externally and internally oriented information⁴⁰. Beyond emotional intensity, we found that semantic similarity to external context supported the retention of internal experiences, both at local and global levels. This suggests that recall may be facilitated both by immediate associative cues from adjacent details^72,73 and by broader contextual cues arising from integration into the overall narrative or event schema^52,74. Notably, the effects of emotion and semantic connection extended beyond internal experiences to external ones, suggesting that they reflect general principles supporting the recall of memory details. In contrast, the number of words describing a memory detail, used as a proxy for informational richness, predicted the retention of external but not internal experiences. This may be because externally oriented details are likely more diverse, encompassing perceptual, spatial, and contextual features, so greater elaboration provides multiple retrieval cues that strengthen retention⁵⁰. By comparison, internal experiences typically revolve around a core thought or feeling and are less influenced by the amount of verbal elaboration.

Our results showed that memories containing more internal thoughts and feelings were perceived as more important. Because this relationship is correlational, at least two explanations are possible. One possibility is that events of greater personal significance naturally elicit stronger internal experiences during encoding¹, which in turn are more likely to be recalled due to the memory-enhancing effects of emotion⁷⁵. Alternatively, recalling more internal experiences may enhance the phenomenological qualities of the memory at retrieval, which are associated with perceived importance⁵⁶. In contrast, memories dominated by external experiences were rated as less important. This may reflect the intrinsic negative relationship between the proportions of internally and externally oriented details within a narrative, or it may be because external experiences often convey factual or perceptual information that, while informative, is less directly tied to personal meaning. Additionally, we found that semantically precise reinstatement of internal experiences was associated with increased perceived importance over time, even as memories overall tended to decline in importance due to fading and reconstruction based on current concerns^33,76. Precise retelling of internal experiences may reinstate the original appraisals and emotions of the event⁷⁷, reengaging the sense of self-relevance that initially made the event important. More broadly, our findings suggest that internal experiences may serve as a key bridge between event and self, capturing not only what happened but also its personal meaning—consistent with the view that autobiographical memory supports self-coherence^3,78,79. These findings also align with perspectives showing that people make meaning from life events by constructing narratives around their memories^38,80,81, with thoughts and emotions playing a significant role in this process^82,83.

Limitations

The present work should be interpreted in light of several methodological considerations. First, participants were asked to recall a memorable event, which may have biased the sample toward emotionally salient experiences²¹, whereas the everyday events that constitute much of daily life are often affectively neutral²⁸. A promising direction would be to combine naturalistic narrative methods with random experience sampling^21,84 to capture internal experiences in narratives of ordinary, real-world events. Second, in narrated free recall, it is difficult to determine whether the absence of certain details reflects genuine forgetting or a choice not to report them. Because memory recall is often a communicative act⁸⁵, participants may omit details they perceive as redundant, peripheral to the task goal, or irrelevant to a potential audience^86,87. Details omitted due to retrieval neglect can also re-emerge when appropriate cues are provided²³. Indeed, details missing from an initial recall frequently appeared in later recalls, as shown in prior studies^30,63 and in the present work. Future research could address this issue by incorporating detailed probes or prompts following narrated free recall. Third, also related to the communicative function of memory recall, the quality and characteristics of autobiographical narratives may vary depending on the social context in real life^86–88. For example, people may share more thoughts and feelings when recalling memories for close others than they do in experimental contexts such as the one used in the present study. Fourth, because the Hippocorpus dataset^43,45 did not include participants’ subjective ratings of individual memory details, we relied on text analysis to characterize them. However, emotion ratings based solely on the text of each individual detail cannot account for the broader context and may not accurately reflect the emotions actually experienced by participants. Moreover, certain phenomenological properties known to predict recall, such as vividness and sense of reliving^23,28, can only be assessed by participants themselves. Collecting such ratings for individual details could therefore provide a more precise basis for predicting retention. Finally, although we grouped thoughts and feelings into a single category consistent with standard autobiographical memory coding schemes³⁹, they may involve distinct processes. Future research could investigate how the properties and retention of memories for thoughts differ from those for emotions, potentially uncovering distinct mechanisms for each.

Conclusion

In sum, our study demonstrates that naturally occurring thoughts and feelings are fragile but influential elements in autobiographical narratives. While these internal experiences were prone to omission, their retention was supported by emotional intensity and stronger connections to other elements of the memory. Recalling such internal experiences was also linked to higher perceived significance of the memory, with the semantic stability of these experiences over time acting as a buffer against the natural decline in significance. By applying recent advances in text and narrative analysis^34,46 to a typically overlooked type of memory—memories of real-world internal experiences—our study broadens the scope of research on long-term memory in humans. These findings may also have implications for clinical contexts, given that the ways people recall their thoughts and emotions are often associated with mental health and well-being^89–91.

Author contributions

H.L. conceived and designed the research. C.K. preprocessed the data. H.S., M. Zhang, and H.L. analyzed the data. H.S., M. Zhang, C.K., and H.L. wrote the original paper. H.S., M. Zhang, M. Zhao, and H.L. reviewed and edited the paper.

Peer review

Peer review information

Communications Psychology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Jesse Rissman and Troby Ka-Yan Lui. A peer review file is available.

Data availability

The original Hippocorpus dataset⁴⁵ is publicly available on the Microsoft website (https://www.microsoft.com/en-us/download/details.aspx?id=105291). A preprocessed version of the subset used in the current study is publicly available on the Open Science Framework (https://osf.io/a65bg).

Code availability

The analyses in this study were conducted using publicly available Python and R packages. For text similarity analysis, we used a pretrained model (all-mpnet-base-v2) implemented in Sentence Transformers (v2.2.0). For sentiment analysis, we used the VADER sentiment analysis tool implemented in the NLTK library (v3.8.1). We used the lme4 (v1.1-35.5) package to perform mixed-effects logistic regression, psych (v2.5.6) to compute Williams tests, cocor (v1.1.4) to obtain confidence intervals and Cohen’s q for correlation comparisons, irr (v0.84.1) to estimate Fleiss’s κ, and MBESS (v4.9.42) to compute confidence intervals for ANOVA effect sizes in R. For other statistical tests, we used Pingouin (v0.5.1), Scipy (v1.12.0), and Scikit-learn (v1.2.2) in Python. Analysis scripts are publicly available on the Open Science Framework (https://osf.io/a65bg).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s44271-026-00425-8.