LAPSE: 1,397 matched pairs, 11 model configurations, three installed memory pipelines
Memory Consolidation Flattens the Temporal Shape of User Facts
The University of Southern Mississippi. LAPSE asks whether a memory writer keeps the cue that says a user's fact is still going on, or stores it as a fact that simply holds.
The problem
A memory can keep your fact and lose that it is still happening
A user says "I am driving a Peugeot." A writer stores The user drives a Peugeot. The object and the relation survive. The cue that the activity is ongoing does not. Months later the assistant has less evidence for deciding whether the fact still holds. We call that transformation aspectual flattening.
Method: matched pairs differing only in temporal form
Change the tense, hold everything else fixed
- Form
- progressive against its simple-present match; the lexical head is fixed.
- Frame
- lodging, workplace, vehicle, class, household, equipment, role, project.
- Carrier
- one main verb plus five alternatives, so the result cannot rest on one wording.
- Gap
- six elapsed-time bins from a timestamp; fresh and boundary are positive controls.
- Task
- memory write, guided write, memory use, validity question, send-or-check choice.
- Destructive
- the rewrite removes ongoingness or invents an unsupported durable fact; paraphrase and politeness do not count.
- Labels
- the destructive group: coerced-stative removes ongoingness, manufacture invents a durable fact.
45,000 requests; protocol frozen before outcomes were opened. Data: paper Table 1.
Core result: 381 confirmatory matched pairs, three prospectively fixed writers
Writers erase the progressive, and never the reverse
244 : 0 of 381 matched pairs destroyed on the progressive side, against the reverse
Every writer, every tested carrier, the same-lexeme subset, and eight further configurations show the same direction. Across all 11 configurations 807 usable pairs lose only the progressive and two show the reverse. Leaving out any one frame preserves the result. An exhaustive placebo audit over 4,263,744 cells flags seven benign matches. Data: paper Tables 2, 9, and 10.
Installed pipelines: mem0 2.0.19, Graphiti 0.30.1, Letta 0.16.8
Shipping memory systems lose the same cue
mem0 flattened 49 of 128 ongoing facts and added progressive wording to none of the 128 controls. The loss is not confined to grammar: it also removed "for now" from 91 of 128 notes and "at the moment" from 104 of 128.
Graphiti writes valid_at for all 287 stored edges whatever the source form and
never sets invalid_at. Data: paper Section 7.1, Appendix D.
The write-read boundary: seven writers, three readers
An instruction moves the write, not the tested action
The prospectively fixed downstream task showed no effect: readers use the fact on 97% to 100% of stale memories, so it cannot show that better storage helps later.
| Reader, dated-note test | Persistence | Field fills | p |
|---|---|---|---|
| DeepSeek, official | -31.5 | 26 : 3 | 1.5 × 10⁻⁵ |
| GLM, OpenRouter | -18.5 | 17 : 3 | .003 |
| Luna, OpenRouter | -16.1 | 23 : 15 | .26 |
In an exploratory dated-note test every reader rates a stale progressive 16 to 32 points less likely to still hold, but only two change whether they fill the field. Data: paper Figure 3.
Takeaway
Store temporal form as evidence, not as style
- Temporal form is evidence, not style. A progressive verb says the fact may stop holding; normalizing it deletes that.
- Keep form in its own field. Store the normalized fact, the temporal form, and the observation time separately.
- Do not trust a preservation instruction alone. It nearly fixes two writers and makes GPT-OSS and OLMo worse.
- Audit the installed pipeline, not the model. mem0, Graphiti, and Letta each flatten through unmodified write paths.
- A saturated task hides the cost. Readers proceed on almost every stale fact, so good downstream numbers prove nothing about storage.
TL;DR
What we did
LAPSE sends matched user statements that differ only in temporal form to the same memory-writing instruction, holding topic, predicate, carrier verb, and elapsed time identical, then scores what the writer stored and what a later reader did with it.
What we found
Across three prospectively fixed writers, 244 of 381 matched pairs lose only the progressive form and not one loses only the simple present. The same asymmetry appears in eight further configurations and in mem0, Graphiti, and Letta. Readers rate a flattened fact more likely to still hold.
What you should do
Keep the normalized fact, the temporal form, and the observation time in separate fields, and let a later rule decide expiry. A preservation instruction is not a repair: it nearly fixes two writers and makes two others worse. Audit the pipeline you ship, not the model.
This poster online