MoM: Memory of Memory

Bowen Qin; Yao Lu

cs.CLSep 7, 2026

MoM: Memory of Memory

Authors: Bowen Qin, Yao Lu

Organizations: National University of Singapore

Abstract

For a long-horizon LLM agent, the memory question is not what was once recorded but what currently holds. Most designs answer it only indirectly: every interaction is stored, and the present is reconstructed at query time by retrieving and reconciling records, so stale values re-enter and the same conflicts are re-litigated. Committing the current value at write time avoids this, but existing write-time (CRUD) memories overwrite, so a wrong update is unrecoverable and prior state is lost. We take the missing combination---commit on arrival while retaining what is displaced---and formalize it as Memory of Memory (MoM): memory tracks not only content but the provenance, status, and history of its own entries. We instantiate MoM as Provenant Memory (P-Mem), a typed provenance graph whose active frontier exposes one current value per resolved key while displaced values are retained as provenance; typed operations decide whether a new observation supports, supersedes, contests, rejects, revokes, or resolves an existing value. P-Mem's decisive gain is validity rather than accuracy: its turn-level read matches the strongest retrieval memory in accuracy at $\sim$ 4 $\times$ fewer read tokens---a retrieval-granularity effect---while graph-guided turn pruning cuts the knowledge-update stale-answer rate (19.4% $\rightarrow$ 10.9%); on revision chains it stays at 100% where query-time reading collapses to 25%, and, because displaced values are retained rather than overwritten, it recovers committed errors a CRUD memory cannot (100% vs.\ 0%).

Figures & tables

Figure 1: One knowledge update across three memory designs. (a) The archive returns the stale value; (b) CRUD memory cannot undo a mistaken update; (c) MoM reads the current value and restores it on retraction.

Design family (systems)	Current value settled	Displaced value	Characteristic failure
Archival: long-context, RAG / note-style, value-aware retrieval (MemRL), tiering (MemGPT / MemoryOS)	query time	retained, not demoted	staleness re-entry; conflicts re-litigated, sometimes inconsistently
CRUD memory (Mem0-style banks)	on arrival	erased (overwritten / deleted)	mis-updates irreversible; history unanswerable
Bitemporal KG (Zep / Graphiti)	on arrival	retained (valid-time intervals)	timestamp-arbitrated; no open conflict, no reversal
MoM (ours)	on arrival	retained as provenance	residual: wrong until revisited (§ 4.2 )

Table 1: Memory designs along two axes: when the current value is settled (on arrival vs. query time) and whether the displaced value is retained or erased. MoM commits on arrival yet retains what it displaces as typed provenance; bitemporal KGs share that corner but arbitrate by timestamp, leaving conflict implicit and commits irreversible.

Figure 2: P-Mem overview. A typed temporal provenance graph turns passive interaction records into an evidence-grounded active frontier of currently valid values, while preserving displaced values as provenance.

Annotation	Head after	Other version(s)
accept	new value (active)	superseded
reject	new value (active)	contradicted
contest	incumbent (contested)	incoming kept as alternative
revoke	predecessor, else empty	revoked (tombstoned)
resolve	promoted alternative (active)	others contradicted
noop	unchanged	unchanged

Table 2: How each typed annotation (§ 3.2 ) lowers onto version-status changes. Head after is the cell’s current value once applied; every other version is retained, retyped as listed.

	Accuracy $\uparrow$		Stale rate $\downarrow$
	oracle	_s	oracle	_s
MemoryOS	66.7	60.3	22.6	19.4
P-Mem, turn read	75.6	69.2	19.6	23.6
P-Mem, turn read $+$ prune	76.9	65.4	5.4	10.9

Table 3: Knowledge-update subset: accuracy and stale-answer rate vs. MemoryOS retrieval on both haystack regimes

	Overall	superseded	aggregation	single
	(500)	(75)	(92)	(333)
No memory	4.0	2.7	0.0	5.4
Long-context	27.0	40.0	5.4	30.0
Mem0	39.4	37.3	35.9	40.8
MemoryOS	45.8	53.3	19.6	51.4
HippoRAG-v2	50.8	58.7	33.7	53.8

Table 4: All 500 LongMemEval-s questions, accuracy by susceptibility class.

	Context tokens $\downarrow$	Accuracy $\uparrow$	Acc. per 1k tok $\uparrow$
Long-context	$\sim$ 106k	27.0	0.25
MemoryOS	$\sim$ 11k	45.8	4.2
P-Mem (turn read)	$\sim$ 3k	52.6	17.5

Table 5: Token efficiency on LongMemEval-s (mean context tokens per query). P-Mem reads

\sim

\times

fewer tokens than MemoryOS retrieval and yields

\sim

\times

more accuracy per token.

	Accuracy (%)					Context (chars)
$L$	Commit	Raw (top-10)	Raw $+$ recon.	Raw $+$ $[t]$ tags	Raw (oracle order)	commit / raw
1	100.0	100.0	100.0	100.0	100.0	30 / 591
2	100.0	72.9	94.3	97.9	100.0	30 / 621
4	100.0	42.9	59.3	80.0	98.6	30 / 640
8	100.0	25.0	40.0	64.3	94.3	30 / 641

Table 6: Current-value accuracy as revisions accumulate (

n{=}140

per

L

, 5 interleaved objects, top-10). Commit stays flat; the raw read must reassemble order and collapses; reconstruction and

[t{=}i]

tags recover part; the arrival-order oracle recovers 94–100%.

	Single-hop acc. $\uparrow$			Multi-hop acc. $\uparrow$			Stale rate (sh) $\downarrow$
	6k	64k	262k	6k	64k	262k	6k	64k	262k
Long-context (full stream)	32%	31%	0%	2%	6%	0%	79.7%	–	–
MemoryOS retrieval (chunks)	41%	28%	16%	8%	2%	0%	39.1%	15.9%	16.9%
Mem0 (CRUD memory)	36%	45%	30%	5%	4%	2%	43.8%	46.4%	59.2%
HippoRAG-v2 (graph RAG)	69%	47%	39%	11%	3%	6%	35.9%	24.6%	22.5%
Graphiti/Zep (bitemporal KG)	15%	32%	12%	1%	1%	4%	18.8%	10.1%	15.5%

Table 7: MemoryAgentBench FactConsolidation at three history lengths (

n{=}100

per cell; stale rate on identifiable chains,

n{=}64

–

71

). Dashes: long-context stale omitted where accuracy collapsed.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

	Recovered $\uparrow$		Wrong	Lost	$v_{1}$ in store
	$d{=}0$	$d{=}32$	( $d{=}32$ )
Commit $+$ retention	100%	100%	0%	0%	100%
Commit, no retention	0%	0%	0%	100%	0%
Raw top-10 (archival)	97.9%	92.0%	6.6%	1.5%	100%
Graphiti/Zep	0%	0%	93.8%	6.2%	100%

Appendix

Table 8: Reversibility under retraction (

n{=}140

, outcomes over committed cases only): recovered

=v_{1}

, wrong

=v_{2}

, lost

=

neither. The two commit arms differ only in retention. Graphiti/Zep commits only 16 errors at

d{=}32

, a small base for its row there.

Gate	Overall $\uparrow$	Agg.	Single	Superseded	Stale $\downarrow$	Pruned
Ungated raw read	52.6	29.3	56.8	62.7	27.4	0
Prune always	45.6	15.2	51.4	57.3	12.9	100
Class-oracle gate	51.8	29.3	56.8	57.3	12.9	15.0
Intent gate (4B)	51.2	29.3	55.6	58.7	25.8	17.2
Question-only router (4B)	47.8	22.8	52.0	60.0	16.1	68.8
Question-only router (35B)	51.8	29.3	55.3	64.0	17.7	22.2

Appendix

Table 9: Gating the provenance prune on all 500 LongMemEval-s questions (%). Per-class accuracy over the judge-assigned classes; the stale rate is over the 62 stale-scoreable superseded questions of this run, a different subset from the 78-question knowledge-update set of Table 3 . Pruned = fraction of questions the gate routes to the pruned read.

	Accuracy $\uparrow$	Stale $\downarrow$	Facts/cell
Turn read, raw	69.2	24.1	—
Turn read + provenance prune	65.4	9.3	—
Materialized state, 4B writer	52.6	33.3	1.17
35B extractor	61.5	20.4	1.12
35B extractor, merge 0.70	57.7	13.0	1.76

Appendix

Table 10: Write-time materialization on the _s knowledge-update subset (

n{=}78

; stale rate over the 54 stale-scoreable updates of this judging pass). All arms share the 4B reader; write-time arms read the full materialized state. Facts/cell measures whether supersession chains form (1.0 = every fact its own cell).

	4B writer	35B extractor
Extraction LLM calls	47.5	47.5
Calls per committed fact	0.15	0.18
Write tokens (prompt $+$ completion)	157k	146k
Ingestion wall time	13.1 s	10.8 s
Committed facts	325	258
Materialized cells	278	231

Appendix

Table 11: Write-side cost per history on the _s knowledge-update subset (78 histories, 47.5 sessions each), measured from per-request logs.

LLM Agent Memory

Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.

See all 376 papers

PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents
Oct 5, 2026Yiqi Wang, Jiaqi Liu, Jiaqi Zhang +4Data Provenance Memory Agent Benchmarks
University of Southern Queensland · Southern University of Science and Technology · Jiangsu University +2
CALM: Conflict-Aware Lifecycle Memory Governance for Long-Horizon Language Agents
Oct 28, 2026Chang Li, Xianhao Ou, Xinyan Ma +7Persistent Memory for Language Models Agent Memory Governance
Harbin Institute of Technology · Meituan
Whose Memory Is It? Scope-Aware Commit Rules for Long-Term LLM Memory
Oct 6, 2026Hongyu Gu, Xinchang LiLong-Term Conversational Memory Agent Memory Management
University of Science and Technology of China Hefei, China

Oct 5, 2026J/K move · Enter open · S save

PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents

Yiqi Wang, Jiaqi Liu, Jiaqi Zhang, Zhangkai Wu+3

University of Southern Queensland · Southern University of Science and Technology · Jiangsu University+2

Data Provenance Memory Agent Benchmarks Graph-Based Agent Memory Agent Memory Management

Open paper AskPDF arXiv

LLM agents rely on long-term memory to retain and reuse information when performing tasks over long horizons. Existing methods provide limited support for handling memories that become outdated as new observations or domain evidence arrive. Such outdated memories may remain semantically relevant, continue to affect dependent records, and retain value as historical evidence. This calls for two capabilities: dependency tracking to identify downstream effects and historical preservation to retain useful past records. We propose Provenance-Aware Cascading Memory Invalidation (PACMI), a framework that represents memories and new evidence in a provenance graph with typed dependency edges. PACMI assigns records to a four-state validity lattice, propagates validity changes to dependent memories, and uses the resulting states for retrieval and stale-premise detection. We also introduce a diagnostic benchmark with 100 cases and 300 queries across five domains. The evaluation separates node, context-, and answer-level performance. PACMI achieves the highest final-answer accuracy on this benchmark, and its paired difference from the strongest baseline is significant under an exact McNemar test. The premise checker achieves perfect precision, recall, and F 1 on the controlled query distribution. Cascading propagation primarily improves memorystate correctness: removing it increases final-answer errors from 3 to 11, but the paired difference does not reach the 0.05 significance threshold. Code and data will be made publicly available.

Abstract

LLM Agent Memory

Explore similar work

PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents

CALM: Conflict-Aware Lifecycle Memory Governance for Long-Horizon Language Agents

Whose Memory Is It? Scope-Aware Commit Rules for Long-Term LLM Memory