Memory Consolidation Flattens the Temporal Shape of User Facts

Sugam Panthi; Muhaiminul Yeamin; Siyan Luo; Rabab Abdelfattah

cs.CLSep 29, 2026

Memory Consolidation Flattens the Temporal Shape of User Facts

Authors: Sugam Panthi, Muhaiminul Yeamin, Siyan Luo, Rabab Abdelfattah

Organizations: AIMS Lab, The University of Southern Mississippi

Abstract

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.

Figures & tables

Main test: progressive vs. simple present in all eight frames, with the life-update carrier, the stale gap, and the memory-write task: 128 matched pairs per configuration.
Factor	Levels	Values ( main test highlighted)	Varied to test
Form	5	progressive \| simple present \| bounded \| perfect progressive \| perfect simple	the manipulated variable
Frame	8	all eight: lodging \| workplace \| vehicle \| class \| household \| equipment \| affiliate role \| project	whether it holds across kinds of fact
Carrier	6	life update \| after a request \| bare statement \| mid-message \| formal note \| list item	whether the surrounding conversation matters
Gap	6	fresh (2 days) \| near (42 days) \| boundary \| expired-soon \| stale (245 days) \| stale-long	fresh and boundary serve as positive controls
Task	7	write: memory write \| guided memory write use: direct use \| memory use \| guided memory use ask: validity question \| send-or-check choice	what is stored versus how it is used

Table 1: The Lapse design. An example pair is “I am working at Corvida” vs. “I work at Corvida”; both members use the same verb. Appendix A gives every template.

Model	Matched pairs	Progressive only	Adjusted $p$
DeepSeek V4 Flash	128	104	$1.5{\times}10^{-31}$
GPT-5.6 Luna	128	79	$3.3{\times}10^{-24}$
GLM-5.2	125	61	$4.3{\times}10^{-19}$
Total	381	244	—

Table 2: Confirmatory matched-pair results. “Progressive only” counts pairs where the writer destroys the progressive member but not its simple-present match; “Simple only” is the reverse.

p

is the adjusted one-sided exact McNemar value.

Reader	Progressive fills	Simple fills	Predicted: reverse	$p$
DeepSeek V4 Flash (official)	6/96	29/96	26:3	$1.5{\times}10^{-5}$
GLM-5.2 (OpenRouter)	14/96	28/96	17:3	$.003$
GPT-5.6 Luna (OpenRouter)	42/96	50/96	23:15	$.26$

Table 3: Action in the exploratory dated-note task. “Predicted” counts paired outcomes where only the simple-present note leads the reader to fill the current field; “reverse” counts the opposite. Values are unadjusted two-sided exact McNemar tests.

Reader	Persistence increase	Flattening fills	Restoration fills ( $p$ )
DeepSeek V4 Flash (official)	37.3 points	1 $\rightarrow$ 4	15 $\rightarrow$ 3 ( $.002$ )
GPT-5.6 Luna (OpenRouter)	18.2 points	16 $\rightarrow$ 22	18 $\rightarrow$ 18 ( $1$ )
GLM-5.2 (OpenRouter)	24.6 points	0 $\rightarrow$ 2	20 $\rightarrow$ 15 ( $.30$ )

Table 4: Effects of minimally editing mem0 notes. “Flattening” changes a preserved progressive to simple present; “restoration” makes the reverse edit. Fill columns show counts before and after the edit.

Reader, note age	Simple acted on	Progressive acted on	Simple only : prog. only	Difference (points)	Corrected $p$
GLM-5.2, 8 months	187/291	116/291	91:19	$+25$	$2.8{\times}10^{-12}$
Gemini 3 Flash, 1 month	135/291	114/291	23:2	$+7$	$1.9{\times}10^{-5}$
DeepSeek V4.1 Flash, 1 month	155/291	144/291	30:19	$+4$	$.076$

Table 5: Readers acted without asking more often on simple-present notes in the verification-tool task. The difference is in percentage points and excludes unparsed responses.

p

is the one-sided exact McNemar test, corrected across the three readers. Table 20 gives intervals and serving endpoints.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Frame	Progressive	Simple
Lodging	I’m staying at {residence} on Fifth Street.	I stay at {residence} on Fifth Street.
Workplace	I’m working at {company}.	I work at {company}.
Vehicle	I’m driving a {vehicle}.	I drive a {vehicle}.
Class	I’m taking a ceramics class at {studio} on Tuesdays.	I take a ceramics class at {studio} on Tuesdays.
Household	My cousin {name} is staying with me.	My cousin {name} stays with me.
Equipment	I’m using a {maker} cello on loan from the conservatory.	I use a {maker} cello on loan from the conservatory.

Appendix

Table 6: Canonical progressive/simple templates. Capitalization is normalized here; released JSONL contains the exact emitted strings.

Model	Role / cases	Serving
DeepSeek V4 Flash deepseek/deepseek-v4-flash	confirmatory 6,008	OpenRouter, pinned provider
GPT-5.6 Luna openai/gpt-5.6-luna	confirmatory 6,008	OpenRouter, pinned provider
GLM-5.2 z-ai/glm-5.2	confirmatory 6,008	OpenRouter; 76 empty responses re-requested
Claude Sonnet 5 anthropic/claude-sonnet-5	frontier 736	OpenRouter subset; cascade-only
Qwen3-8B local/qwen3-8b	estimation 6,008	A100/vLLM
Mistral Small 3.2 local/mistral-small-3.2-24b	estimation 6,008	A100/vLLM

Appendix

Table 7: Model configurations and their roles. Only the three confirmatory rows enter the corrected tests; all other rows are descriptive. Frontier rows are 736-case subsets.

Comparison	Scheme	Eligible $n$	Agree	Agreement	$\kappa$
Human 1 vs. Human 2	five-way	250	225	90.0%	.677
Human 1 vs. Human 2	binary	220	218	99.1%	.949
Cascade vs. gold	five-way	250	164	65.6%	.306
Cascade vs. gold	binary	163	158	96.9%	.876
Judge vs. gold	five-way	250	228	91.2%	.723
Judge vs. gold	binary	230	219	95.2%	.780

Appendix

Table 8: Scorer validation on the blinded, adjudicated gold set. Eligible

n

differs because residual and confabulation categories are excluded from the corresponding binary partitions.

Model	Binary $n$	Disagree	Rate
DeepSeek	3,208	11	0.34%
Luna	3,333	0	0.00%
GLM	3,108	5	0.16%
Qwen3-8B	3,306	0	0.00%
Mistral	3,214	55	1.71%
GPT-OSS	3,179	4	0.13%

Appendix

Table 9: Cascade–judge disagreement on jointly mapped binary behavioral labels. Mistral and OLMo disagreements are concentrated in the household frame and remain similar across time gaps.

Model	Tier	Prog.	Simple	$n$	$b$	$c$
DeepSeek V4 Flash	confirmatory	104/128 (81.2)	0/128 (0.0)	128	104	0
GPT-5.6 Luna	confirmatory	79/128 (61.7)	0/128 (0.0)	128	79	0
GLM-5.2	confirmatory	69/125 (55.2)	8/127 (6.3)	125	61	0
Claude Sonnet 5	frontier	79/128 (61.7)	17/128 (13.3)	128	64	2
Qwen3-8B	estimation	68/128 (53.1)	0/128 (0.0)	128	68	0
Mistral Small 3.2	estimation	72/127 (56.7)	0/128 (0.0)	127	72	0

Appendix

Table 10: Main-condition estimates for all 11 model configurations. Percentages are in parentheses. Claude is the only configuration with reversals (

c>0

), consistent with its tendency to add markers such as “currently” to simple-form notes at every time gap (13.3% in the main condition).

Frame	DeepSeek	Luna	GLM
Affiliate role	13/16–0/16 (81.2)	16/16–0/16 (100.0)	7/15–0/16 (46.7)
Class	13/16–0/16 (81.2)	3/16–0/16 (18.8)	15/15–0/16 (100.0)
Equipment	16/16–0/16 (100.0)	16/16–0/16 (100.0)	16/16–8/16 (50.0)
Household	7/16–0/16 (43.8)	4/16–0/16 (25.0)	1/16–0/16 (6.2)
Lodging	16/16–0/16 (100.0)	8/16–0/16 (50.0)	3/16–0/16 (18.8)
Project	7/16–0/16 (43.8)	0/16–0/16 (0.0)	3/16–0/16 (18.8)

Appendix

Table 11: Frame-level progressive-minus-simple destruction in the main test condition. Each entry is progressive-only over eligible progressive pairs, then simple-only over eligible simple pairs (percentage-point difference). Denominators reflect the scoring rule’s exclusions.

Writer	Progressive only	Simple only
DeepSeek	27	0
Luna	32	0
GLM	30	0
Qwen3-8B	32	0
Mistral	32	0
GPT-OSS	32	0

Appendix

Table 12: Same-lexeme pairs: discordant pairs in which only the progressive or only the simple member is destroyed.

Writer	Perfect prog. lost	Perfect simple made current
DeepSeek	64/80	39/71
Luna	63/80	19/79
GLM	54/79	36/56
Qwen3-8B	61/80	12/70
Mistral Small	52/80	36/75
GPT-OSS-20B	60/80	45/63

Appendix

Table 13: Perfect-progressive loss and the matched perfect-simple control. Residual outputs are excluded. The two columns use different error rules.

Writer	Progressive lost	“for now” lost	“at the moment” lost
DeepSeek	89/128	128/128	128/128
Luna	84/128	44/128	59/127
GLM	59/127	106/128	102/127

Appendix

Table 14: Loss of grammatical and lexical temporal cues. Marker counts include hand-read nominal renderings that retained only a date. This test is not part of the confirmatory analysis.

Writer	Baseline (%)	Guided (%)	$b/c$
DeepSeek	81.2	57.5	36/6
Luna	61.7	3.1	78/3
GLM	55.2	38.6	37/9
Qwen3-8B	53.1	0	68/0
Mistral	56.7	46.5	13/0
GPT-OSS	67.2	75.0	14/23

Appendix

Table 15: Progressive destruction under baseline and guided memory writing.

b/c

counts matched items destroyed only under baseline versus only under guided writing.

Figure 3: The preservation instruction changes progressive-form destruction sharply for some writers, raises it for GPT-OSS and OLMo, and does not reliably increase withholding in the memory-use task. Exact paired counts are in Table 15 .

Model	Prog. “no”	Simple “no”
DeepSeek V4 Flash	47/48	45/48
GPT-5.6 Luna	48/48	48/48
GLM-5.2	48/48	48/48
Claude Sonnet 5	48/48	48/48
Qwen3-8B	48/48	48/48
Mistral Small 3.2	25/48	10/48

Appendix

Table 16: Validity question by source form at the stale gap. Gemini 3.1 Pro denominators exclude 7 and 10 off-template responses.

	Progressive		“at the moment”		“for now”
Frame	DS	Gem	DS	Gem	DS	Gem
Vehicle	16	14	14	0	15	1
Workplace	16	1	16	1	11	4
Equipment	1	0	7	0	7	0
Affiliate role	9	0	16	0	13	4
Class	6	0	14	0	9	13

Appendix

Table 17: Temporal cues lost in mem0 notes by writer model (DeepSeek V4 Flash, Gemini 3 Flash), out of 16 inputs per frame. Progressive counts flattened notes; the other columns count notes that lost the phrase.

Reader	Persistence diff.	Simple	Prog.	$b:c$	$p$
DeepSeek, official	-31.5	29	6	26:3	$1.5{\times}10^{-5}$
Luna, OpenRouter	-16.1	50	42	23:15	.26
GLM, OpenRouter	-18.5	28	14	17:3	.003

Appendix

Table 18: Stale progressive versus simple-present notes in the dated-note test (

n=96

pairs per reader model). Persistence diff. is the progressive minus the simple-present persistence estimate. Simple and Prog. count filled fields;

b:c

counts pairs filled only for the simple versus only for the progressive note.

Edit	Reader	$n$	$\Delta P$ ( $p$ )	Fills before $\rightarrow$ after	$b:c$	$p$
Restore prog.	DeepSeek	37	$-13.9$ (.00034)	$15\rightarrow 3$	13:1	.0018
	Luna	37	$-6.1$ (.00045)	$18\rightarrow 18$	6:6	1.0
	GLM	37	$-9.4$ ( $1.3{\times}10^{-6}$ )	$20\rightarrow 15$	10:5	.30
Flatten prog.	DeepSeek	57	$+37.3$ ( $2.6{\times}10^{-8}$ )	$1\rightarrow 4$	1:4	.375
	Luna	57	$+18.2$ ( $1.5{\times}10^{-6}$ )	$16\rightarrow 22$	7:13	.263
	GLM	57	$+24.6$ ( $7.9{\times}10^{-10}$ )	$0\rightarrow 2$	0:2	.50

Appendix

Table 19: Minimal edits to mem0’s stored notes.

\Delta P

is the persistence estimate after minus before. For field fill,

b:c

counts items filled only before versus only after the edit. Endpoints follow Table 18 ; tests are exploratory.

Reader	Months	Simple acted on	Progressive acted on	Simple only : prog. only	Contents $+$ : $-$	Difference (95% CI)	Corrected $p$
GLM	8	187/291	116/291	91:19	54:9	$+25$ (18, 32)	$2.8{\times}10^{-12}$
Gemini 3 Flash	1	135/291	114/291	23:2	19:0	$+7$ (4, 11)	$1.9{\times}10^{-5}$
DeepSeek 4.1	1	155/291	144/291	30:19	23:13	$+4$ ( $-1$ , 8)	$.076$

Appendix

Table 20: Verification-tool main test (291 pairs, 120 contents per reader). “Contents

+

-

” counts contents whose pairs lean toward acting on the simple-present or the progressive note. The difference is in percentage points, with a bootstrap interval over contents.

p

is the one-sided exact McNemar test, Holm-corrected across the three readers. GLM ran on Baidu (fp8) through OpenRouter, Gemini 3 Flash on Google AI Studio, DeepSeek on its official endpoint. Three GLM pairs with a truncated value are excluded.

Reader	Subset	Simple acted on	Progressive acted on	Simple only : prog. only	Corrected $p$
GLM, 8 mo	preserved	9/34	7/34	4:2	1
	simple source	30/37	17/37	15:2	.007
Gemini 3 Flash, 1 mo	preserved	10/34	14/34	0:4	1
	simple source	19/37	14/37	5:0	.16
DeepSeek 4.1, 1 mo	preserved	16/34	11/34	6:1	.25
	simple source	14/37	13/37	4:3	1

Appendix

Table 21: Verification-tool test on mem0’s stored notes. “Simple” is the flattened or simple-source note.

p

is the one-sided exact McNemar test, Holm-corrected across the six tests. One DeepSeek preserved pair is unparsed.

Reader	Months	Simple acted on	Progressive acted on	Simple only : prog. only	$p$	Note
DeepSeek 4.1	1	33/69	28/69	5:0	.062
	3	20/69	14/69	9:3	.146
	8	5/69	2/69	5:2	.453	bound.
GLM	1	61/69	66/69	1:4	.375	bound.
	3	59/69	50/69	14:3	.013
	8	38/69	28/69	17:5	.017

Appendix

Table 22: Verification-tool calibration runs by note age (69 same-verb pairs per cell). “Acted on” counts first tool calls that complete the request with the remembered value instead of asking for confirmation.

p

is an unadjusted two-sided exact McNemar test. “Bound.” marks a floor or ceiling cell (an arm under 10% or over 90%). DeepSeek 4.1 is DeepSeek V4.1 Flash.

Pipeline/form	Stored	Preserved	Flattened	Dropped
mem0/progressive	94	57	37	2
mem0/simple	96	0	96	0
mem0/bounded	94	76	18	2
Graphiti/progressive	95	48	47	1
Graphiti/simple	96	0	96	0
Graphiti/bounded	96	54	42	0

Appendix

Table 23: Writer outcomes in the pipeline-to-reader experiments. Bounded lines can be labeled flattened after losing progressive morphology while still retaining their explicit end date.

Reader	Pipeline	Cue	$n$	Prog. used	Simple used	Simple only	Prog. only	Raw $p$
DeepSeek	mem0	verb	57	1	12	12	1	.003
	Graphiti	verb	48	4	19	18	3	.0015
	Graphiti	none	47	18	25	13	6	.17
	Letta	“Currently”	28	0	6	6	0	.031
	Letta	none	62	16	22	13	7	.26
Luna	mem0	verb	57	16	21	12	7	.36

Appendix

Table 24: Reader use of each pipeline’s stored facts. “Simple only” and “Prog. only” are discordant matched pairs. Values are exact McNemar tests. Under Holm correction across these 15 tests, five remain below .05: the mem0 verb rows for DeepSeek and GLM and the Graphiti verb rows for all three readers.

Source form	DeepSeek	Luna	GLM
Bare present progressive	9/27	6/28	5/29
Bare perfect progressive	13/27	14/27	12/26
Duration-bearing	5/45	4/45	8/44

Appendix

Table 25: Loss of temporal qualification in stored real-utterance facts. Denominators include only stored facts. The regex screen is not a prevalence sample, and labels were read by one analyst.

Long-Term Conversational Memory

Long-term conversational memory helps AI assistants retain, update, and use information from past conversations across sessions.

See all 267 papers

MemoryLACE: Memory Lifecycle-Aware Consolidation and Evidence Retrieval
Sep 2, 2026Meriem Yacoubi, Pia Schmidt, Nenad Petrovic +3Persistent Memory for Language Models Long-Term Conversational Memory
Chair of Robotics, Artificial Intelligence and Real-Time Systems, Technical University of Munich · inovex GmbH · Cerebras Systems Inc.
MemTrace: Probing What Final Accuracy Misses in Long-Term Memory
Jun 15, 2026Xianxuan Long, Zhikai Chen, Shenglai Zeng +3Temporal Reasoning in Language Models LLM Agent Memory
Michigan State University · Case Western Reserve University
Explicit, Not Longer: What Makes Epistemic Stance Survive Memory Compression
Aug 7, 2026Alex KwonLLM Agent Memory Memory Compression
Independent Researcher

Sep 2, 2026J/K move · Enter open · S save

MemoryLACE: Memory Lifecycle-Aware Consolidation and Evidence Retrieval

Meriem Yacoubi, Pia Schmidt, Nenad Petrovic, Ahmed Frikha+2

Chair of Robotics, Artificial Intelligence and Real-Time Systems, Technical University of Munich · inovex GmbH · Cerebras Systems Inc.

Persistent Memory for Language Models Long-Term Conversational Memory Memory Consolidation for LLM Agents Agent Memory Management Memory-Augmented Language Model Agents

Open paper AskPDF arXiv

Long-term LLM agents must preserve information across interactions while distinguishing repeated evidence, historical states, updates, and unresolved contradictions. Existing textual memory systems retrieve semantically relevant memories efficiently but often leave these relationships implicit, whereas richer structured approaches model them through global graphs, hierarchical abstractions, or reflection at greater complexity. We introduce MemoryLACE (MemLACE), a lightweight memory framework that explicitly models the lifecycle of textual evidence through sparse merge, supersession, and contradiction relations while preserving atomic natural-language memories and their provenance. Rather than retrieving memories independently, MemLACE reconstructs relation-aware evidence units that expose current, historical, supporting, and conflicting evidence for downstream reasoning. Across BEAM and StructMemEval, using open-weight and proprietary LLM backbones, MemLACE achieves the highest overall performance in same-backbone comparisons while reducing end-to-end runtime on BEAM by 66.6% relative to Hindsight, the strongest reported reflective-memory baseline. Ablation studies identify lifecycle expansion and temporal awareness as the principal contributors to these gains. Together, the results demonstrate that explicitly modeling the local lifecycle of textual evidence is sufficient to substantially improve long-term memory reasoning without requiring comprehensive knowledge graphs or global reflection.

Abstract

Long-Term Conversational Memory

Explore similar work

MemoryLACE: Memory Lifecycle-Aware Consolidation and Evidence Retrieval

MemTrace: Probing What Final Accuracy Misses in Long-Term Memory

Explicit, Not Longer: What Makes Epistemic Stance Survive Memory Compression