Finding positive meaning in memories of negative events adaptively updates memory

Nature Communications 12: 6601 (2021), DOI 10.1038/s41467-021-26906-4, PMC8593143, CC BY 4.0. Columbia, Rutgers–Newark and Icahn School of Medicine at Mount Sinai. Data and code: osf.io/jtgfk (not read). Provenance: papers/carry_on/speer2021_positive_meaning_memory.provenance.json.

What was read

All 565 lines of text converted from the PMC open-access XML by pmc2txt.py, with the self-check passed:

• abstract, results, discussion and methods of all four experiments;
• the captions of Figures 1–5;
• data and code statements, and references.

Not read: the figures (images, captions only), the Supplementary Information (exclusions by group, baseline analyses, anhedonia, the monetary-reward cohort, whole-brain maps), the peer review file and the source data.

Question

When people find positive meaning in a negative autobiographical memory after recalling it, does the memory itself change? Does it later come back feeling more positive, with different content? Does the change last, and does it depend on reactivation, as reconsolidation would predict?

Method

• Experiment 1 (N = 102 of 131 recruited).
  • People write about 12 negative memories and rate "How does this make you feel in the present moment?" on a scale from −5 to +5.
  • They then elaborate each memory in one of four groups: positive meaning, negative aspects, neutral details, or a distraction task.
  • One week later they recall and rate the memories again. Two blind raters score the positivity of the content and how much it differs.
• Experiment 2 (N = 91 of 128, online): positive meaning against natural re-recall, with tests at 1 week and 2 months.
• Experiment 3 (N = 72 of 104): a reconsolidation design.
  • Everyone elaborates half their memories positively and half naturally. Groups differ in timing:
    • Delayed test: mental recall, 10 minutes, positive elaboration, test 24 hours later;
    • Immediate test: the same, but tested 1 hour later;
    • No reminder: elaboration 24 hours after recall, test 24 hours later.
• Experiment 4 (N = 32, fMRI): the delayed-test design with scans 24 hours apart.
  • Representational similarity analysis in hippocampus, ventral striatum and ventromedial prefrontal cortex (VMPFC).
  • A two-month behavioural follow-up, with n = 18.
• Exclusions. Only memories that people reported as "the same memory" at each recall were analysed.
  • Memories were also excluded if general, not negative, or not elaborated as instructed: 21.2% of memories in Experiment 1 and 23.1% of trials in Experiment 4.

Results

• Feeling changes most after positive meaning (Experiment 1).

  • The groups differed in feeling change after a week: F(3, 98) = 4.08, p = .009.
  • The positive group improved more than each other group (d = 0.64–0.81). The others did not differ from each other.
  • Group differences in content were not significant (positivity p = .116; dissimilarity reported p = .326).
  • Within the positive group only, feeling change tracked more positive content (r = .52) and more changed content (r = .43).

• It lasts two months (Experiment 2). Against natural recall, the positive group was higher at both 1 week and 2 months on:

  • positive feeling (d = 0.50 and 0.53);
  • positive content (d = 0.48 and 0.43);
  • content dissimilarity (d = 0.70 and 0.64).

  All groups also became more positive over time. The authors link this to the fading affect bias.

• What changes in the content (Experiment 2).

  • Later recollections kept most original details: 74.6% and 72.9% in the positive group, against 82.6% and 80.3% in the control group.
  • They absorbed about 10% of the positive elaboration.
  • They added more new positive details (12.2% and 14.1%, against 8.1% and 10.2%). New negative details did not differ.
  • Original details made up 66.7% against 79.8% of recollections at 1 week, and 64.7% against 78.4% at 2 months.

• Timing matters (Experiment 3). Only the delayed-test group showed more positive feeling for positive than for control memories (against immediate test d = 0.74; against no reminder d = 0.66). The authors read this as reconsolidation-dependent updating.

• Neural (Experiment 4).

  • Positive memories felt better than control memories after 24 hours (d = 0.91) and after 2 months (n = 18, d = 0.76).
  • Pattern dissimilarity across recalls tracked increased positivity more for positive than for control memories in hippocampus (d = 0.42) and ventral striatum (d = 0.43), not in VMPFC.

• It does not work for everyone. Anhedonia was associated with less updating (SI, not read).

Limits

• Healthy young student samples. Memories are self-selected cues from a list, and "the same memory" is self-reported.
• The outcome is a single bipolar feeling scale; positive and negative emotion are not separated.
• Content is scored by raters on 10-point scales; the text cannot be shared.
• Repeated measures are analysed as independent observations (inferred from the degrees of freedom).
  • Experiment 2 uses F(1, 178) with 91 people at two time points.
  • Experiment 3 uses F(·, 137) with 72 people in two conditions.
  • The significance levels are likely overstated.
• The key Experiment 3 test, the condition × group interaction, is not significant (p = .181). The reconsolidation claim rests on follow-up t-tests.
• The two-month fMRI follow-up retained 18 of 32 participants.
• Inconsistencies found:
  • Experiment 1's content-dissimilarity ANOVA, F(3, 98) = 1.53, is reported with p = .326. My computation gives p = .21. Either way it is not significant.
  • Experiment 2's feeling ANOVA gives η² = .023 for time (F = 11.82) and .061 for group (F = 4.43). Both effects have 1 df and share one error term, so η² must scale with F. The values fit only if swapped (.061/.023 = 2.65 against 11.82/4.43 = 2.67; my computation).
  • Experiment 3's condition (2) × group (3) interaction is reported as F(1, 137). A 2 × 3 interaction has 2 numerator df.
  • Within the positive group (n = 26), correlations are reported as r25 in one place and r24 in another. Other groups use n − 2.
  • Race and ethnicity percentages sum to 107.9% (Experiment 1) and 112.6% (Experiment 4); how Hispanic was counted is not stated.
• Checked:
  • Exclusion arithmetic in all four experiments: 131 → 102, 128 → 91, 104 → 72, 40 → 32.
  • Group sizes and the men counts.
  • The p-values for Experiment 1's feeling ANOVA and for the Experiment 2 and 3 main effects.

What it means for Kurisutina

• Q3: recall is where memories change, and what is thought at recall gets written in.
  • Recalling a negative memory and then dwelling on its positive side produced later recollections that felt better and contained new positive details, for at least two months, with most original details kept.
  • For a replica, every retrieval of a stored experience is a possible rewrite. What the replica thinks while retrieving it, including what the base model adds, can be folded into the next version.
• This is the mechanism by which drift can enter memory (inferred).
  • Suppose a replica's memory store is rewritten after each retrieval: a consolidated summary, or Mem0-style UPDATE.
  • The base model's framing during retrieval then enters the memory, much as positive elaboration did here.
  • Human memory does the same thing, but with the person's own framing. The design question is whose reappraisal gets written in, not whether any does.
  • A Q3 test (proposal): retrieve stored episodes repeatedly under neutral instructions; measure after each cycle:
    • drift in affect and content toward the model's typical framing;
    • retention of original details, against the human 73–83% after two months.
• Timing and consolidation windows (inferred). Updating needed a reminder, then elaboration, then a delay. For a replica, the analogue is which retrievals trigger a rewrite.
  • Rewriting on every read is maximally labile.
  • Humans are labile only after reactivation, and their changes consolidate.
  • A replica could rewrite only when retrieval is followed by elaboration, logging old and new versions (proposal). This agrees with the brief's "never overwrite".
• Person-level moderation again. Anhedonia reduced updating, and the fading affect bias is reduced in dysphoria. How much a person's memories soften with time and reappraisal is a person trait. A replica must soften at that person's rate, not the default human one.
• Same-memory control. The paper analyses only memories that people confirm are "the same memory". Q3 scoring should likewise separate a changed version of the same memory from retrieval of a different memory.

Cross-references

• summaries/carry_on/ritchie2015_fading_affect_pancultural.md: negative affect fades faster over time.
• summaries/carry_on/stjacques2013_reactivation.md: reactivation-induced updating of real-world memories.
• summaries/carry_on/schacter2011_adaptive_distortion.md: adaptive functions of memory distortion.
• summaries/carry_on/haehner_perception_stability.md: event appraisals change over a year.
• summaries/carry_on/chhikara2025_mem0.md: an LLM memory that overwrites on update.