Memory distortion: an adaptive perspective

Trends in Cognitive Sciences 15(10): 467–474, DOI 10.1016/j.tics.2011.08.004. Read as the NIH author manuscript (PMC3183109; accepted, not copyedited). Harvard. A short review and opinion article, no new data. Guerin and St. Jacques contributed equally. Provenance: papers/carry_on/schacter2011_adaptive_distortion.provenance.json.

What was read

All 1,107 lines of the text derived from the PMC XML: the abstract, the main text (introduction, three domains, concluding remarks), Boxes 1–4, the captions of Figure 1 and Box Figure I, the acknowledgements and all 100 references. The figures themselves are images and were not inspected; their captions state the results. There is no supplement.

Question

Are memory distortions evidence of a flawed memory system, or side effects of processes that make memory useful?

Argument and evidence

The authors review three kinds of distortion and link each to an adaptive process. "Adaptive" is used loosely, meaning "beneficial"; they do not claim adaptations in the strict evolutionary sense.

• Imagination inflation. Imagining a novel event raises confidence that it happened, produces claims of having done or seen it, or produces full false recollections.
  • Mazzoni and Memon: imagining raised belief and false memories more than reading about the event did. For an impossible event (a nurse removing a skin sample from the little finger, a procedure not performed in the UK), false memories were four times as likely after imagination as after exposure, and about 30% of those who imagined it developed a false memory.
  • Imagining an act produces about as many false memories of having done it as a doctored video suggesting one did.
  • Imagery regions (precuneus, inferior parietal, fusiform) are more active at encoding for items later falsely remembered as seen.
  • Adaptive account: the constructive episodic simulation hypothesis. Remembering the past and imagining the future use one "core network" (medial prefrontal and parietal, hippocampus, lateral temporal and parietal). Episodic memory serves the future by recombining stored elements into simulations; the cost is miscombining imagination with memory.
• Gist-based and associative errors.
  • In the DRM paradigm, a list such as candy, sour, sugar, … leads people to falsely recall or recognise the unstudied "sweet". Roediger and McDermott found false recognition around 80%.
  • Evidence that these errors come from a healthy system:
    • amnesic patients make fewer of them;
    • true and false recognition share brain regions, but only for gist-related lures, not for unrelated false alarms;
    • left ventrolateral prefrontal activity at encoding, linked to semantic elaboration and later true recognition, also predicts later gist-based false recognition;
    • false recall of a DRM lure primed solutions to remote-associate problems, and DRM susceptibility is predicted by convergent thinking, a component of creativity.
  • Adaptive account: gist and associations support generalisation, abstraction and compact storage.
• Post-event misinformation. Wrong information given after an event is later endorsed as part of it; true and false memories share encoding regions.
  • Edelson et al.: after a group viewing, fabricated "co-observer" answers raised errors on questions people had answered correctly days earlier. Some errors persisted after debriefing that the answers were random. Only those persistent errors engaged the hippocampus and medial temporal lobe at the time, read as real updating of the memory.
  • Hupbach et al.: a reminder of object set 1 before learning set 2 led to set-2 intrusions when set 1 was recalled 48 h later. The effect was asymmetric, so it is not simple source confusion; it is read as reactivation opening the memory to new information (reconsolidation).
  • Adaptive account: a memory that updates flexibly with relevant new information pays for it when source accuracy matters, as in eyewitness settings.
• Also noted, briefly.
  • People remember their past in ways that flatter their present self-view, which can help well-being (Ross and Wilson).
  • Whether negative or positive events are distorted more is contested.
  • McKay and Dennett consider only positive illusions to be strict adaptations.
• Box 1, retrieval conditions. Details can be stored even when gist errors occur.
  • After studying 2,896 pictures for 3 s each, people distinguished fine details (Brady et al.).
  • In Guerin et al. (Experiment 2, N = 30, all comparisons p < .001), a similar lure was falsely recognised often when shown with unrelated items, and still often next to another similar item, even though eye tracking showed attention to the relevant details. It was rarely chosen when shown beside the studied target.
  • Conclusion: gist errors here come from a lack of access to stored detail at retrieval, not from loss of detail.
• Box 2, contextual associations.
  • After about 10 minutes in an office, people falsely recalled typical office objects that were not there.
  • Encoding activity in retrosplenial, medial prefrontal and lateral parietal cortex predicted later false recognition of contextually related new objects.
  • "Context frames" support prediction, at the price of such errors.
• Box 3, sleep. Sleep increased DRM false recall in some studies and had no effect or reduced it in others. The authors propose two effects: stabilisation, which favours detail, and integration with existing knowledge, which favours gist and false memory.
• Box 4, open questions. Other adaptive processes; individual differences (creativity and associative errors); criteria for strict adaptations versus by-products; the roles of encoding, storage and retrieval.

Limits

• A short opinion and review piece. The evidence is selected to fit the adaptive view, and the "adaptive" claim is explicitly loose.
• Nearly all paradigms are laboratory ones (word lists, pictures, short films). Real autobiographical memory over months and years is touched only through the self-enhancement and misinformation work.
• Several cited effects are contested in the text itself: sleep, the role of plausibility in imagination inflation, and positive versus negative events.
• A metadata error in the PMC reference list: reference 26 (Mazzoni & Memon, Psychol. Sci. 2003) carries the DOI 10.1046/j.1432-1327.1999.00020.x, which belongs to a different journal (European Journal of Biochemistry).

What it means for Kurisutina

• It supports the candidate prediction for question 3. Perfect retention is itself a source of divergence. The person's memory of a new experience is transformed by processes that make memory useful: gist extraction, schema-driven completion, updating from later information, and self-flattering reconstruction. A replica that keeps a verbatim record of the same experience will differ from the person in predictable directions:

  • it keeps details the person lost or cannot access;
  • it lacks the gist and context intrusions the person has;
  • it does not fold later information into the memory the way the person does, or it folds it in too readily (Park et al. 2023's "memory hacking").

• Carrying on as the person means modelling their memory transformation, not only storing more. This is a design proposal, in the provenance terms of brief 5.3:

  • keep the verbatim record as the source layer, for audit;
  • drive behaviour from a held by the person layer that is allowed to change the way the person's memory changes;
  • fit the change on that person's own measured forgetting and distortion. Whether individual differences in these distortions are stable enough to fit is open; Box 4 asks the same question.

• A test for question 3 follows directly. Person and replica take in the same new material (a story, a short film, a staged conversation). Both are then tested after a delay with:

  • free recall;
  • recognition with gist lures and context lures;
  • a misinformation phase, including socially attributed misinformation as in Edelson et al.

  Score whether the replica's error profile matches the person's: which details are lost, which lures are accepted, which misinformation sticks. Accuracy alone is not the measure. A replica that remembers better than its person fails this test.

• Elicitation design rules (brief 6.2 and 6.3):

  • Asking a person to imagine an unconfirmed event can create a memory of it: about 30% for an impossible event. Imagination prompts about unverified episodes should not be used, or should be logged as a perturbation.
  • A reminder followed by new information can merge the two (Hupbach). An interviewer who cues an old episode and then supplies related material risks writing that material into it. This supports the perturbation audit.
  • What a person "has" depends on the probe: details inaccessible to free recall can be available when the original item is re-presented (Box 1). This matches the brief's use of recognition probes with re-presented material for the provenance arm (Slotnick & Schacter, brief [11]).

• The replica's own simulations are an imagination-inflation risk. Park et al.'s agents write reflections and plans into the same memory stream as perceptions. Imagined and remembered content then share one store, as they share one network in people. Tagging every replica-generated simulation as such (the source field) is the engineering counterpart of reality monitoring, and it needs to be checked, not assumed.

• Social conformity is a drift channel. Misinformation from co-observers persisted even after debriefing. A replica's susceptibility to what its interlocutors assert should match its person's, neither the chat model's agreeableness (Li et al. 2024; Park et al. 2023) nor zero.

Cross-references

• summaries/carry_on/park2023_generative_agents.md: memory stream, reflections, memory hacking.
• summaries/carry_on/li2024_persona_drift.md: the agent takes on its interlocutor's instruction.
• summaries/11_slotnick_schacter2004.md (brief [11]): the sensory signature separating true from false recognition.
• summaries/memory/roediger1995_false_memories.md: the DRM paradigm.
• summaries/memory/lindsay1990_source_suggestions.md, summaries/memory/johnson1993_source_monitoring.md, summaries/memory/lindsay1989_eyewitness_source.md: misinformation and source monitoring.
• summaries/memory/nader2000_reconsolidation.md, summaries/memory/schiller2010_reconsolidation.md, summaries/memory/chalkia2020_reconsolidation_replication.md: reconsolidation. The last is a registered replication that found no persistent reactivation-extinction effect on human fear memory; weigh it before leaning on reconsolidation as the mechanism of updating.