Forgetting the unforgettable through conversation: socially shared retrieval-induced forgetting of September 11 memories

Psychological Science 20(5): 627–633, DOI 10.1111/j.1467-9280.2009.02343.x. New School for Social Research and Lehman College, CUNY. © Association for Psychological Science. Read as the published PDF posted publicly on a Squarespace site (publisher typesetting; site owner not identified). Provenance: papers/carry_on/coman2009_forgetting_conversation.provenance.json.

What was read

All 711 lines of pdftotext output in reading order (the layout mode interleaves this two-column page): abstract, both experiments, general discussion, references and Table 1. Figure 1, the only figure, was rendered and its bars read (values below are plot readings).

Question

Does recounting some memories in conversation make related unmentioned memories harder to access, for the speaker and for a listener? Does this hold when speaker and listener hold different memories of the same kind of event?

Method

• Material. Each participant's own September 11 memories, collected with a standard questionnaire.
  • Answers were grouped into eight category sets of four (six experimental, two filler), such as time and location.
  • Final test: a participant-specific yes/no recognition test. Correct probes came from the person's own answers; lures were plausible answers taken from other participants.
• Experiment 1 (n = 20, New York residents on 9/11).
  • After a 25-minute distraction, a structured interview practised 6 questions from three category sets. That makes three item types:
    • Rp+: practised;
    • Rp−: unpractised, from the same sets;
    • Nrp: from unpractised sets.
  • Recognition test after 5 minutes.
• Experiment 2 (22 unrelated participants in pairs).
  • Day 1: the questionnaire. Day 2: pairs recounted their 9/11 experiences to each other, freely, for about 7 minutes, then took the recognition test.
  • The experimenter nudged topics in 82% of conversations.
  • Coders classified each item by practice type and by the participant's role: speaker (mentioned their own answer) or listener (the partner mentioned a related answer of their own). Category sets where both partners mentioned the item (9%) were dropped.
• Outcome. Reaction times for hits, as "retrieval-induced inaccessibility", because a single conversation was not expected to cause outright forgetting of well-rehearsed memories. Errors were analysed as well.

Results

• Experiment 1: retrieval-induced forgetting within one person. Reaction times differed by practice type (F(2, 38) = 8.55, ηp² = .31). Rp+ was faster than Nrp, and Nrp faster than Rp−.

  Item type	RT, ms (fig)	Error rate (fig)	Contrast	t(19)	d
  Rp+	≈ 1,150	≈ .02	Rp+ vs Nrp	2.73	0.60
  Rp−	≈ 1,690	≈ .14	Nrp vs Rp−	2.30	0.56
  Nrp	≈ 1,360	≈ .07

  Errors showed the same direction as trends only.

• Experiment 2: the same for speakers and for listeners, from a free conversation.

  Role	Rp+ (fig)	Rp− (fig)	Nrp (fig)	Key test
  Speaker	≈ 1,380 ms	≈ 1,670 ms	≈ 1,350 ms	Nrp faster than Rp−: t(18) = 2.63, d = 0.56. Omnibus a trend: F(2, 36) = 2.64
  Listener	≈ 1,230 ms	≈ 1,610 ms	≈ 1,300 ms	F(2, 30) = 7.45, ηp² = .33. Nrp faster than Rp−: t(16) = 2.61, d = 0.52

  • For listeners, the rehearsal effect was a trend (t(17) = 2.09).
  • Error rates did not differ by item type in Experiment 2.
  • The listeners' slowed items were their own memories, not the speaker's. Hearing someone else recount a related part of a different experience made one's own related details slower to access.

• The effect is about 300 ms (Rp− against Nrp) in every panel (fig). There was little or no rehearsal benefit for speakers in Experiment 2 (Rp+ ≈ Nrp).

Limits

• Samples. 20 and 22 people. Of the 22, only 16–19 enter each Experiment 2 analysis; the exclusions are not explained (probably empty cells after coding; my inference).
• Outcome. Only accessibility, by reaction time, minutes after practice. Errors showed no effect in Experiment 2. So this is not lasting forgetting; the authors say so.
• Design. Massed practice; one recognition format; the experimenter steered the conversations; the delay between conversation and test in Experiment 2 is not stated.
• Inconsistencies found (minor; my recomputation from the reported F and t values):
  • Three of the four "p <" bounds are slightly too low:
    • F(2, 38) = 5.63 gives p = .0072, reported "p < .007";
    • F(2, 36) = 2.64 gives p = .085, reported "p < .08";
    • F(2, 30) = 7.45 gives p = .0024, reported "p < .002".
  • The error-rate ηp² for F(2, 38) = 5.63 computes to .229, reported ".22".
  • The d values are not the paired d_z (t/√n: 0.61, 0.51, 0.41, 0.40, 0.60, 0.63), and the method is not stated. Two nearly equal t values (1.82 and 1.79) carry d = 0.50 and 0.63.
• Checked:
  • ηp² for F(2, 38) = 8.55 (.310), F(2, 36) = 2.64 (.128) and F(2, 30) = 7.45 (.332).
  • The Experiment 1 item counts: 6 Rp+, 6 Rp−, 12 Nrp and 16 distractors, over 24 + 16 = 40 probes.

What it means for Kurisutina

• Q3: remembering selectively reshapes what stays accessible, even for vivid memories.

  • A person who retells part of a memory makes the related untold parts harder to reach, and so does a listener about their own experiences (verified).
  • A replica whose memory store only adds weight to what is retrieved has half of this. It lacks the suppression of related, unretrieved details. This is the same point as the 2016 follow-up (verified in both papers; the replica claim is inferred).

• The listener effect points both ways (inferred).

  • A replica talking with the person's family would, by what it chooses to recount, shift which of their own related memories stay accessible. That is an effect on people outside the slot, and worth stating in any real deployment.
  • A replica that "listens" to a user recount related events could, if built to be human-like, lose access to details of its own related episodes. Whether that is desired is a design decision, not a default.

• Test design (proposal, as in the 2016 summary). The Rp+/Rp−/Nrp scheme gives a ready Q3 probe:

  1. Build category sets from the replica's own episodic record.
  2. Have it recount a subset in conversation.
  3. Compare accessibility of related and unrelated untold items, by retrieval rank or latency in the store, or by recall accuracy.

  People show Rp− slower than Nrp by about 300 ms in minutes. A store without competition will show none.

Cross-references

• summaries/carry_on/coman2016_mnemonic_convergence.md: the network-level follow-up (convergence, retrieval-induced forgetting), read earlier.
• summaries/carry_on/hirst2009_911_memory.md: long-term retention of 9/11 memories.
• summaries/carry_on/talarico2003_flashbulb_confidence.md: consistency and confidence of 9/11 memories.
• summaries/carry_on/zhong2023_memorybank.md: retrieval-strengthened memory without competition.