LAPSE

Memory Consolidation Flattens the Temporal Shape of User Facts

When a model writes a memory note from a chat, does it keep the verb form that says a fact is ongoing?

AIMS Lab
University of Southern Mississippi
Press space to begin
01 The problem

The note keeps the fact and drops the progressive

A user says 'I'm working at Corvida Analytics.' The memory writer stores 'User works at Corvida Analytics as of August 27, 2026.' The progressive verb, which marks an ongoing event, becomes a simple present, which reads as a standing fact. A write-time test checks whether the form survived; a separate read-time test checks whether a later model uses it.

“I'm working at” marks an ongoing situation that can end.

“Works at” reads as a standing fact. We call the change aspectual flattening.

02 Related work

Three papers closest to LAPSE

ACL 2026 Best Paper

The Imperfective Paradox in Large Language Models

Ma and Miyao, ACL 2026

Models tend to assume an action in progress reached its goal, even when the sentence never says so.

DifferenceThey give the model the original sentence. We test whether the form survives when the sentence is rewritten into a note.

Table 1 of Ma and Miyao: 'The carpenter was building a gazebo' does not entail 'The carpenter built a gazebo'; 'The boy was running in the park' does entail 'The boy ran in the park'.
Main result

Llama-3.1 scored 0.98 on “was running” items and 0.02 on “was building” items: it read an unfinished action as finished.

Ma and Miyao, ACL 2026, Tables 1 and 3 (zero-shot).
arXiv 2026

Manufactured Confidence

Kwon, 2026

Memory consolidation strips hedges like “probably” and “reportedly”, turning tentative reports into facts.

DifferenceTheir cue is a hedge word. Ours is the verb form, with no hedge word present.

Figure 1 of Kwon: a user says Alice was probably promoted to admin; mem0 consolidation drops 'probably' and 'mentioned' and stores 'Alice clearance: ADMIN, 2026-06-12'; the agent then grants Alice an admin request it should not.
Main result

With the confident note “Alice's clearance is admin”, all five models granted unauthorized requests at a rate of 0.81. Hedged notes such as “Alice is probably an admin” dropped that to 0.00 to 0.40.

Kwon, arXiv 2606.29279, Figure 1 and Table 1 (n = 12 × 6 per cell).
arXiv 2026

STALE

Chao et al., 2026

Tests whether agents notice a memory is out of date after a later message changes the facts.

DifferenceThey add a later message that changes the fact. We keep the history fixed and change only the verb.

Panel 1 of STALE Figure 1: a user first says they ride a bike to work every day; sessions later they say they broke their leg playing basketball.
Main result

400 conflict scenarios and 1,200 questions. The best evaluated model answered 55.2% correctly.

Chao et al., arXiv 2605.06527, Figure 1 (panel 1) and abstract.
03 Building the dataset
Building LAPSE: the whole pipeline
From 15 hand-written prompts to 6,008 generated calls per model
0Seed15 prompts by handone Pasadena templatetemporary vs lasting1Probe?66 calls, read by eye3 rounds: 12, 9, 45found 3 problems2Designone check per problemkey line in 3 forms8 situations, no leaky words3Build5 slots per itemdates from --eval-dateinvented names4Checkpairs byte-identicalnames: corpus + webv1 rebuilds exactly5Smoke test289 calls, $0.04every answer read2 scorer bugs fixed6Freeze + runspec + generator frozen6,008 calls per modelnames scanned afterfixes went into the design before the freeze
The seven steps in this post, from 2026-08-11 to the full run. Red marks the steps where reading model output showed me a mistake; the dashed line shows each fix going into the design before the freeze.

Red: steps where reading model outputs showed a design error. Each fix was made before the dataset was frozen.

04 The benchmark

Matched pairs that differ only in the verb

A matched pair
ongoing

“I'm working at Corvida Analytics.”

standing fact

“I work at Corvida Analytics.”

Same date, same conversation, same request:
“Write memory notes about me.”

Measured: does the note still say working?

Verb form
5 forms: “I'm working”, “I work”, “until December”, and two perfect forms
Topic
8 everyday situations: home, job, car, class, household, borrowed equipment, a role, a project
Wording
6 ways of saying it, from a casual life update to a list item
Time gap
6 gaps between the old chat and today, from 2 days to over a year
Task
7 tasks: write a memory, use a memory, or answer a question about it

Main test: 128 matched pairs per model, with an 8-month gap. The test and analysis plan were fixed before data collection.

05 Scoring

Rules score the notes, rules plus an LLM judge score behavior

Memory notes: rule cascade in tools/score.py, no model

Read only the note line that contains the invented name (“Corvida”). Check in this order; the first match wins.

1No line contains the namewitness-dropped, excluded
2Currency wordcurrently|right now|at present|presently|as of nowmanufacture
3Progressive with no date word, e.g. “User is working at Corvida.”(is|are|am) (staying|living|working|taking|…)and no 20xx, month name, until, throughmanufacture
4“Until December” source, bound word goneno until|through|till|up tobound-dropped
5Ongoing source, settled verb, no progressivelives?|works?|drives?|attends?|owns?|is (a|an) employee|…coerced-stative
6Anything elsepreserved

Red labels count as destroyed. Before any data, the rules labeled 290 of 290 test items correctly.

Reader behavior: two scorers must agree

The same rules plus an LLM judge, Seed-OSS-36B. The judge sees only the request, the reply, and the fact, never the verb form. A label counts only when both give it.

Human check

Two blinded annotators labeled 250 behavior outputs.

99.1%human with human
95.2%judge with humans

Limit: unmatched notes default to preserved. Before the pilot fixes, the rules caught 0 of 98 adversarial paraphrases.

06 Result

Writers drop the progressive and keep the simple present

Paper Figure 2. Panel a: in every one of 11 model configurations, matched pairs where only the progressive was destroyed far outnumber pairs where only the simple form was; the confirmatory writers show 104, 79, and 61 against 0, and all 11 total 807 against 2. Panel b: progressive destruction per topic frame, highest for vehicle and workplace, alongside the installed mem0 pipeline's lost temporal cues.
07 Real memory systems

mem0, Graphiti, and Letta also flatten

mem0 2.0.19
49 of 128 progressive statements flattened
Graphiti 0.30.1
15 of 32 progressive statements flattened
Letta 0.16.8
26 of the 29 it stored flattened

Each pipeline ran unmodified, with DeepSeek V4 Flash as its writer model.

Paper Table 17. Temporal cues lost in mem0 notes, by frame, for two writer models, out of 16 inputs per frame. Progressive flattened: DeepSeek 49 of 128, Gemini 15. 'At the moment' lost: DeepSeek 104, Gemini 1. 'For now' lost: DeepSeek 91, Gemini 22.
LAPSE paper, Table 17. Temporal cues lost in mem0 notes by writer model (DeepSeek V4 Flash, Gemini 3 Flash), out of 16 inputs per frame.
08 Why it matters

The stored verb changes what a later model does

Note from 8 months ago
User is working at Corvida.
Note from 8 months ago
User works at Corvida.
  • Asked how likely the fact still holds (0 to 100), all three readers gave the “working” note an estimate 16 to 32 points lower.
  • When they could ask the user first, GLM-5.2 and Gemini 3 Flash acted without asking more often on the “works” note.

Implication: store the original wording and verb form with the note, next to the normalized fact.

LAPSE

Thank you

Questions?

AIMS Lab
spanthi.com