PaperlayerView exact revision on arXiv

Source-linked AI summary

Memory Consolidation Flattens the Temporal Shape of User Facts

Sugam Panthi, Muhaiminul Yeamin, Siyan Luo, Rabab Abdelfattah

arXiv:2609.36457v1cs.CL

TL;DR

Memory writing can preserve a user fact while dropping temporal-form evidence that helps later models judge whether it still holds. The paper introduces LAPSE, a benchmark of matched statements that differ only in temporal form, and finds selective flattening and effects on later readers. It concludes that systems making validity decisions should retain temporal form as evidence.

  • Problem

    Memory notes may preserve the fact while losing temporal-form evidence that helps a later reader judge whether it still holds.

    Source: §1 · p. 1

  • Method

    LAPSE uses matched user statements that differ only in temporal form to test what memory writers preserve and whether later readers use the surviving cue.

    Source: §1 · p. 2, §4.1 · p. 3

  • Results

    In 244 of 381 pairs, writers flattened progressive statements while keeping their simple-present matches, with no reversals; the pattern held across tested configurations and installed pipelines, and verb-only edits shifted all three readers’ persistence estimates.

    Source: §1 · p. 2, §9 · p. 9

  • Takeaways & Limitations

    Memory systems making later validity decisions should store temporal form as evidence instead of discarding it.

    Source: §8 · p. 8, §9 · p. 9

    LAPSE uses synthetic, English-only stimuli and omits many pragmatic, discourse, and language-specific ways of signaling that a fact is temporary.

    Source: §10 · p. 9

Abstract

from arXiv · show

Long-term memory systems turn conversations into short stored notes. A note can keep a user fact while losing evidence about whether the fact still holds. For example, "I am driving a Peugeot" can become "The user drives a Peugeot," which drops the cue that the activity is ongoing. We call this aspectual flattening and measure it with LAPSE, a benchmark of matched user statements that differ only in temporal form. We find that memory writers flatten aspect selectively. Three writer models flattened the progressive statement but kept its simple-present match in 244 of 381 pairs, never the reverse. The asymmetry holds in all 11 model configurations tested and in the installed pipelines mem0, Graphiti, and Letta. The lost cue matters to later readers. In exploratory tests, changing only the stored verb shifted all three readers' estimates that a fact still holds. When readers could ask the user before acting, two of three acted without asking more often on flattened notes. Our planned memory-use task could not detect this, because readers there acted on almost every stored fact, even expired ones. Memory writing can thus remove evidence that later models use to decide whether to act.

1 Introduction

The paper asks whether memory writing preserves temporal-form evidence that helps later readers judge whether a fact still holds. It introduces LAPSE to isolate this question and finds selective flattening alongside downstream effects on readers.

1 Introduction

  • A memory note can retain a fact while dropping the cue that its activity is ongoing, leaving a later reader with less evidence to decide whether to check it first.The paper calls this transformation aspectual flattening.Source: §1 · p. 1
  • LAPSE tests matched user statements that differ only in temporal form, with controls for general action tendencies and content that sounds temporary.The benchmark separates whether writers preserve the source cue from whether later readers use the surviving cue.Source: §1 · pp. 1–2
  • The paper names aspectual flattening and releases LAPSE with its stimulus builder, model outputs, labels, and analysis programs.Source: §1 · p. 2
  • In 244 of 381 matched pairs, writers flattened the progressive while keeping its simple-present match, with no reversals; the direction held across all 11 configurations and installed pipelines.Later readers’ persistence estimates also shifted when only the stored verb changed, and some acted without asking more often on flattened notes.Source: §1 · p. 2

2 Related Work

Prior memory systems use temporal metadata and mechanisms, while existing evaluation settings often focus on recall, updates, or changed circumstances. LAPSE instead tests whether writing preserves temporal evidence already present in a user’s wording.

2 Related Work

  • Long-term memory systems use timestamps, recency, decay, graph chronology, consolidation, timelines, or lifecycle policies to manage stored information.These systems reason over stored events, metadata, or relations among memories.Source: §2 · p. 2
  • Long-horizon evaluations cover recall, temporal reasoning, updates, and abstention, while stale-memory studies often supply a later observation or contradiction that changes the world.LAPSE holds the history fixed and changes only the source wording.Source: §2 · p. 2
  • Other memory studies examine lost epistemic hedges or relations between facts; LAPSE tests decision-relevant qualification carried by verb form alone.Its focus is whether the write step removes that cue before a later reader sees the note.Source: §2 · p. 2
  • Aspect diagnostics test whether models infer an event’s endpoint from an available premise, whereas LAPSE tests whether a writer preserves present-progressive evidence for a later reader.Its same-lexeme control addresses concerns about lexical variation and label mapping.Source: §2 · p. 2

3 Temporal Form as Memory Evidence

Temporal form conveys information about how a situation unfolds, distinct from the fact and timestamp a memory may retain. LAPSE treats loss of overt ongoingness or boundedness as destructive when that form bears on whether a fact still holds.

3 Temporal Form as Memory Evidence

  • Progressive form typically presents an event as ongoing, while simple present often presents a standing state; both readings are defeasible in context.For example, “I am working at Corvida” presents the job as ongoing and possibly temporary, while “I work at Corvida” presents it as a standing state.Source: §3 · pp. 2–3
  • Table 1 illustrates matched statements using the same verb, while Figure 1 shows a pipeline replacing progressive “working” with simple-present “works” but retaining a date.Source: Table 1 · p. 3, Figure 1 · p. 3
  • A rewrite is destructive when it removes overt marking of ongoingness or boundedness, or adds an unsupported fact that cancels that marking.Other paraphrase changes can still preserve the source’s temporal meaning.Source: §3 · p. 3
  • A timestamp records when a statement was recorded, while aspect helps express what kind of claim was recorded; they carry different information.Linguistic form is one source of validity evidence, and LAPSE is not designed to rank systems.Source: §3 · p. 3

4 The LAPSE Evaluation

LAPSE uses matched user-statement pairs to vary temporal form while holding other wording and context fixed. Its evaluation separates memory writing from memory use and specifies scoring and confirmatory controls for both.

4.1 Factorial design

  • The factorial design pairs statements with identical topic, main verb, carrier, and elapsed time, varying only temporal form while also testing frames, carriers, gaps, and tasks.The main test uses progressive–simple pairs in a life-update carrier at a 245-day stale gap; fresh and boundary gaps serve as positive controls.Source: §4.1 · pp. 3–4

4.2 Write and read tasks

  • The write task asks a model to store a note, while use tasks give that note to a later model for a validity judgment or send-or-check choice.Guided writing adds an instruction to preserve temporal qualification; send-or-check results were null or degenerate and appear in Appendix C.6.Source: §4.2 · p. 4

4.3 Scoring

  • Write outputs are labeled by a deterministic rule cascade, while behavior labels combine the cascade and an open-weight judge when they agree.Two blinded humans labeled a 250-item gold set; human–human binary agreement was 99.1% (218/220, κ = .949), and judge–gold agreement was 95.2% (219/230, κ = .780).Source: §4.3 · p. 4

4.4 Confirmatory analysis and controls

  • Confirmatory analysis compares progressive-only destruction with the reverse in matched pairs, corrects three tests for multiple comparisons, and checks controls for alternative explanations.A prediction is confirmed when at least two configurations are significant after correction and all three show the predicted direction; other carriers use a sign test.Source: §4.4 · p. 4

5 Consolidation Flattens Temporal Form

Memory writers selectively erase progressive form while usually preserving matched simple-present statements, with the asymmetry persisting across configurations and robustness checks. Flattening varies by frame and instruction, and also appears in real user text.

5 Consolidation Flattens Temporal Form

  • 244 of 381 matched pairs lost only the progressive form, with no reversals in the three confirmatory configurations.The comparison tests progressive-only against simple-only destruction.Source: §5 · p. 4, Table 2 · p. 5
  • Across 11 configurations, writers destroyed only the progressive in 807 usable pairs and only the simple form in two.Both reversals came from Claude Sonnet 5; Gemini 3.1 Pro had 17 unparsed outputs.Source: §5 · p. 5
  • The asymmetry held across four robustness checks, including same-lexeme pairs, alternative carriers, frame omissions, and resampling.Same-lexeme pairs yielded 217 progressive-only destructions and no reversals across seven complete configurations.Source: §5 · p. 5
  • On screened real-user text, writers lost present progressives in 5–9 of 27–29 cases and perfect progressives in 12–14 of 26–27 cases.The inputs came from WildChat, LoCoMo, and Long-MemEval.Source: §5 · p. 5

5.1 What predicts flattening

  • Vehicle and workplace progressives were flattened most often, while household and project progressives were flattened least; no frame reversed the direction.Matched simple-present statements survived in 23 of 24 writer-by-frame combinations; GLM-5.2 destroyed 8 of 16 equipment-frame controls.Source: §5.1 · p. 5
  • DeepSeek and GLM also removed “for now” and “at the moment,” while Luna preserved these lexical cues more often than the progressive.Removed cues were usually replaced by a date.Source: §5.1 · p. 5
  • A tense-and-aspect preservation instruction nearly eliminated progressive destruction for Luna and Qwen3-8B but worsened results for GPT-OSS-20B and OLMo-2-32B.It partly helped DeepSeek V4 Flash and had smaller benefits for GLM-5.2 and Mistral Small 3.2.Source: §5.1 · p. 5
  • Qwen3-8B and Qwen3.8-Max differed by only 1.2 percentage points in raw progressive destruction in the main test condition.Overall flattening rates varied by frame, writer, and writing instruction.Source: §5.1 · p. 5

6 Released Memory Systems Also Flatten

Tests of released memory pipelines found progressive flattening, but the cue removed varied by pipeline, writer, and frame.

6 Released Memory Systems Also Flatten

  • With DeepSeek V4 Flash, mem0 flattened 49 of 128 progressive statements, while none of 128 matched simple-present controls gained progressive wording or “currently.”It also removed “for now” from 91 notes and “at the moment” from 104, usually leaving a date.Source: §6 · p. 5
  • For lodging statements, mem0 flattened none of 16 progressives, although the controlled prompt destroyed all 16; mem0 instead removed “at the moment” from all 16 such notes.This illustrates that the pipeline can change which temporal cue is lost.Source: §6 · pp. 5–6, Figure 2 · p. 6
  • Graphiti flattened 15 of 32 progressive inputs, and Letta flattened 26 of the 29 it stored, often as label-style facts.With Gemini 3 Flash, mem0 flattened 15 of 128 progressives, almost all in the vehicle frame, and added “currently” to 16 simple-present equipment notes.Source: §6 · p. 6

7 Readers Judge and Act on the Stored Form

The paper tests whether later readers use temporal form in notes, first in dated-note tasks, then with verb-only edits and choices to act or verify. Results show that verb form can affect persistence estimates and, for some readers and settings, actions; the planned memory-use task was too insensitive to detect temporal evidence.

7.1 Controlled notes

  • In the planned memory-use task, readers used 59–60 of 60 facts with passed end months and used 97–100% of stale facts, leaving withholding at or below 5.5%.Because readers acted on nearly every stored fact, this task could not register weaker temporal evidence.Source: §7.1 · p. 6
  • For stale notes, all three readers rated progressive notes 16–32 points below matched simple-present notes in persistence.Only DeepSeek V4 Flash and GLM-5.2 reliably changed field filling; Luna’s paired comparison was not significant.Source: §7.1 · p. 7

7.2 Notes written by installed pipelines

  • When a pipeline preserved a temporal cue in its verb, readers filled the current-information field less often than for matched simple-present notes in all nine comparisons.After correction, five of six kept-verb comparisons remained significant; none of the three “Currently” comparisons did.Source: §7.2 · p. 7
  • mem0 date-stamped 23 of 37 flattened progressive notes but only six matched simple notes, and DeepSeek V4 Flash and GLM-5.2 distinguished them by date.The installed-pipeline notes were read eight months after their write date.Source: §7.2 · p. 7
  • Readers showed no reliable difference on Graphiti’s 47 flattened facts or Letta’s 62 cue-free labels; these null tests do not establish equivalence.Graphiti’s valid_at field equaled write time regardless of source form, while showing its unset end field made readers treat preserved progressives as more likely to hold.Source: §7.2 · p. 7

7.3 Verb-only edits

  • Flattening 57 progressives preserved by mem0 raised every reader’s persistence estimates, while its effect on field filling was weaker.Restoring progressives in 37 matched notes reduced filled fields for DeepSeek V4 Flash, but not GPT-5.6 Luna or GLM-5.2.Source: §7.3 · pp. 7–8

7.4 Acting without checking

  • With eight-month-old notes, GLM-5.2 acted without asking on the simple-present note alone in 91 pairs, versus 19 progressive-only pairs.The difference was significant after correction and held on 51 facts not used in calibration; the table reports the verification-tool comparison.Source: §7.4 · p. 8, Table 5 · p. 8
  • Restoring the progressive in mem0 notes cut GLM-5.2’s unasked actions from 30 to 17, but this was the only of six pipeline tests to pass multiple-comparison correction.A surviving date almost always made GLM-5.2 ask, so the verb mattered only on notes without one.Source: §7.4 · p. 8

8 Discussion

Aspectual flattening varies across frames and writer models, so aspect should not be treated as uniform. The authors recommend retaining source wording, normalized facts, temporal form, and observation time as separate fields.

8 Discussion

  • Frame-level flattening varies, and which frames lose the cue depends on the writer model.LAPSE cannot determine whether the variation comes from grammar, content priors, or their interaction.Source: §8 · p. 8
  • A useful memory record keeps original wording, normalized fact, temporal form, and observation time in separate fields.The authors caution against treating aspect as an expiration date because it is defeasible.Source: §8 · p. 8
  • Graphiti’s valid_at marks when a fact was learned, not when it began; readers treated an unset invalid_at as strong evidence of continued validity.That unset field could override a preserved verb in the pipeline-to-reader tests.Source: §8 · p. 8

9 Conclusion

The paper introduces LAPSE to measure whether memory writing preserves temporal-form cues and to distinguish writer behavior from later reader behavior. Writers selectively flatten progressive statements, and the stored verb affects later readers’ persistence estimates and actions.

9 Conclusion

  • Across tested configurations, memory writers flattened progressive sources more often than matched simple-present sources; installed pipelines also flattened them, with every tested writer doing so in writer-dependent frames.Source: §9 · pp. 8–9
  • LAPSE measures temporal-cue loss with matched statements and separates what a writer stores from what a later reader does.The benchmark consists of matched user statements differing only in temporal form.Source: §9 · p. 8
  • Changing only the stored verb shifted all three readers’ persistence estimates; flattened notes also led GLM-5.2 and Gemini 3 Flash to act without asking more often.The paper concludes that memory systems making later validity decisions should retain temporal form as evidence.Source: §9 · p. 9

10 Limitations

The study’s evidence is bounded by its synthetic English stimuli, limited system coverage, exploratory downstream tests, and serving variation.

10 Limitations

  • LAPSE uses synthetic, English-only stimuli with neutral vocabulary, omitting many pragmatic, discourse, and language-specific signals of temporariness.Scoring may not transfer to structured or multilingual memories; bounded-form claims rely on three frames with a computed end month.Source: §10 · p. 9
  • Installed-pipeline checks used one version per pipeline and cannot estimate how often flattening occurs in deployed products.mem0 used two writer models; Graphiti and Letta used one each.Source: §10 · p. 9
  • Downstream action tests used synthetic notes, one sample per item at temperature zero, and a prompt specifying when to rely on memory or confirm.They measured acting without asking, not whether an action produced a wrong outcome.Source: §10 · p. 9
  • Identical temperature-zero calls varied by about five persistence points on the vendor endpoint and about three times as much across OpenRouter hosts.Confirmatory configurations did not log the serving host.Source: §10 · p. 9

Ethical Considerations

The paper evaluates memory writing and use with synthetic, controlled materials and reports both methodological checks and results across model configurations, robustness tests, installed pipelines, and exploratory readers.

C.1 All model configurations

  • Across all 11 model configurations in the main condition, Claude alone had reversals (c > 0), consistent with adding markers such as “currently” to simple-form notes at every time gap (13.3%).Source: Table 10 · p. 12

C.3 Same-lexeme and carrier robustness

  • Across seven complete configurations, same-lexeme pairs showed 217 progressive-only destructions and no reversals; all five alternative carriers had the predicted direction for each confirmatory model.The result also held when frames were omitted or resampled.Source: §C.3 · p. 13

C.4 Perfect forms

  • Every complete writer often turned perfect progressives into current statives, while perfect-simple controls were also sometimes rewritten as current.The authors describe these perfect-form results as descriptive.Source: §C.4 · p. 13, Table 13 · p. 13

C.5 Lexical-marker test

  • DeepSeek and GLM usually converted “for now” or “at the moment” to dates, while Luna retained those markers more often than it retained the progressive.The lexical-marker test was not part of the confirmatory analysis.Source: §C.5 · p. 13, Table 14 · p. 13

C.6 Write intervention and downstream task

  • In the memory-use task, withholding was at most 5.5% under either write instruction, and no model-level baseline-versus-guided comparison was significant.Figure 3 plots write-time and downstream withholding rates; the smallest exact p-value was .125.Source: §C.6 · p. 13, Figure 3 · p. 14

C.7 Validity question and accidental-match scan

  • Except for Mistral, configurations answered “no” near-universally for both stale simple and progressive forms, indicating generic caution rather than sensitivity to ongoingness.Mistral’s answers moved with form in the expected direction.Source: §C.7 · p. 14, Table 16 · p. 14

D Installed-Pipeline Details

  • mem0 flattened 49 of 128 progressive facts and removed “for now” from 91 of 128 notes and “at the moment” from 104 of 128.Gemini 3 Flash flattened 15 of 128 progressive statements in the same inputs; the results are summarized in Table 17.Source: D Installed-Pipeline Details · p. 14, Table 17 · p. 15

E Exploratory Reader Details

  • The dated-note reader test found stale progressive-minus-simple persistence contrasts of -29.3, -15.6, and -18.7, while stale simple-present notes were used in only 28–50 of 96 cases.The verb-only minimal-edit test held note format, date, prompt, fact, and response options fixed; Table 19 reports paired results.Source: E Exploratory Reader Details · p. 15, Table 19 · p. 15

E.1 Verification-tool task

The verification-tool task tested whether temporal form shifts later readers’ choice to use a stored value or ask the user. Several pipeline verb-row contrasts survived correction, while other results depended on how pipelines retained temporal cues.

E.1 Verification-tool task

  • The main verification test crossed 120 contents with two or three request wordings, yielding 291 pairs per reader; three candidates could not be tested.The eight-month GLM calibration row on 69 pairs was not significant after correction; the excluded candidates had confirmation, expired-note, or tool-call issues.Source: §E.1 · pp. 15–16
  • GLM’s per-pair splits favored acting on simple-present notes for affiliate role (27:1), class (31:0), workplace (12:4), project (12:7), and vehicle (6:1).Equipment split in the opposite direction (3:6); Gemini never acted on class notes in either form, and DeepSeek’s project notes went the other way (4:10).Source: §E.1 · p. 16

F Pipeline-to-Reader Details

  • Five of 15 reader-use tests remained significant after Holm correction: mem0 verb rows for DeepSeek and GLM, and Graphiti verb rows for all three readers.Table 24 reports these pipeline-to-reader comparisons.Source: Table 24 · p. 17
  • Pipeline reader contrasts depended on retained cues: mem0 moved some cues into date stamps, while Letta retained “Currently” on 28 lines.Graphiti’s predicted null held for 47 cue-absent facts and its predicted contrast for 48 verb-preserved facts; the nulls do not prove equivalence.Source: F Pipeline-to-Reader Details · p. 17
  • Graphiti set valid_at to write time on all 287 witness edges and extracted explicit endpoints from 91 of 96 bounded inputs.It never set invalid_at for a progressive input; endpoint extraction contradicted the prediction fixed before the run.Source: F Pipeline-to-Reader Details · p. 17

G Benchmark Screen

  • Among screened LongMemEval and LoCoMo utterances, authors found no benchmark question answerable only by noticing temporary form.The 80 utterances came from a high-recall search with unknown recall, and the finding reflects the authors’ reading of that selected set.Source: G Benchmark Screen · p. 17

G.1 Writer test on screened utterances

  • The screened-utterance writer test passed 104 life-circumstance statements through three writers, but its regex screen is not a prevalence sample.The labels were read by one analyst; Table 25 reports stored-fact outcomes without supplied cell values here.Source: §G.1 · p. 17, Table 25 · p. 18

H Reproducibility and Release

  • The confirmatory protocol was fixed before data collection but was not registered with a third party; the release includes raw traces and reproducible analysis artifacts.Figures are built only after analysis outputs reproduce committed hashes.Source: H Reproducibility and Release · p. 18

Related papers

Loading 2609.36457v1…