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

Chang Li; Xianhao Ou; Xinyan Ma; Hao Li; Haoyuan Xu; Jiafeng Liang; Xiaohan Wang; Jiajun Chai

cs.AIOct 28, 2026

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

Authors: Chang Li, Xianhao Ou, Xinyan Ma, Hao Li, Haoyuan Xu, Jiafeng Liang, Xiaohan Wang, Jiajun Chai,

Organizations: Harbin Institute of Technology · Meituan

Abstract

A memory can be outdated for present-day reasoning yet indispensable for reconstructing the past. Relevance-based retrieval overlooks this tension: semantic relevance, evidence-grounded validity, and query-specific eligibility are distinct. We propose CALM, a conflict-aware lifecycle memory governance framework that decouples persistent memory preservation from query-time admission. CALM converts heterogeneous records into evidence-grounded memory cells, updates validity and lifecycle states using support, conflict, and dependency signals, and admits query-appropriate evidence through training-free or learned gating informed by temporal applicability. Across LongMemEval-S and LoCoMo with Qwen and GPT backbones, CALM outperforms state-of-the-art baselines, achieving the best overall accuracy. Compared with simple filtering and reranking, query-conditioned admission further improves evidence selection. Conflict-blocking and robustness analyses show CALM suppresses erroneous retrieved memories while retaining useful evidence. Historical-reconstruction experiments demonstrate that preserving superseded records supports accurate earlier-state reasoning without compromising current-state answers.

Figures & tables

Figure 1: Schematic comparison of relevance-only retrieval and CALM. CALM retains earlier records for historical queries while excluding incompatible evidence from current-state reasoning.

Figure 2: Overview of CALM. (a) Evidence-Grounded Memory Adaptation converts heterogeneous backend records into canonical memory cells. (b) Lifecycle-Aware Memory Governance updates their validity and lifecycle states using support, conflict, dependency, and temporal signals. (c) Query-Conditioned Memory Gating admits eligible memories via CALM-ZS or CALM , while retaining non-admitted memories in persistent storage for traceability and future compatible queries.

Panel of Table 1 — Table 1: Main results on LongMemEval. Accuracy (%) by method across question types. SS denotes single-session, Pref. preference, Asst. assistant, K-Upd. knowledge-update, Multi multi-session, and Temp. temporal reasoning. The best and second-best results within each backbone are shown in bold and underlined, respectively.

Qwen3-30B-A3B-Instruct-2507
Method	Overall	S-Hop	M-Hop	Open	Temp.
Mem0	59.09	65.16	54.26	45.83	51.40
MemOS	76.04	81.81	73.76	59.38	67.91
EverMemOS	87.99	92.75	86.17	70.83	82.24
LightMem	72.01	78.60	72.34	57.29	58.88
CALM-ZS	85.78	92.15	84.04	67.71	76.01

Table 2: Main results on LoCoMo.

Variant	EMB $\uparrow$	UMR $\uparrow$	CL $\downarrow$
No-gating	0.00	100.00	100.00
CALM-ZS	79.32	67.76	21.68
CALM	81.30	72.06	20.00

Table 3: Conflict-blocking performance.

Variant	Overall Acc. ( $\uparrow$ )
Stage-wise ablations
w/o Stage I	69.80
w/o Stage II	75.60
w/o Stage III	74.00
Simple filtering and reranking
Temporal-Filter	74.40

Table 4: Stage-wise ablations and simple alternatives on LongMemEval-S with Qwen3-30B-A3B-Instruct-2507.

Figure 3: Performance of memory management strategies in reconstructing current and historical states.

Variant	Overall $\uparrow$	K-Upd. $\uparrow$	Temp. $\uparrow$
w/o Conflict	73.00	89.74	62.41
w/o Temporal	72.60	88.46	60.15
w/o Validity	72.40	87.18	63.16
w/o Compatibility	74.20	89.74	63.91
w/o Risk	75.20	89.74	64.66
CALM	79.60	91.03	68.42

Table 5: Lifecycle and gating ablation results.

Figure 4: Accuracy across memory-token budgets.

Method	Acc. $\uparrow$	Tok. $\downarrow$	R/G $\downarrow$	Ans. $\downarrow$	Total $\downarrow$
EverMemOS	66.40	2775	46.05/–	14.39	61.65
CALM-ZS	76.60	3850	6.02 / 0.40	1.44	12.00
CALM	79.60	3971	6.02 / 1.03	1.49	12.84

Table 6: Results on query-time efficiency.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Evidence source	Weight $\nu$
LongMemEval-S	Raw dialogue span	0.74
	Rule-derived span	0.82
	Session anchor	0.62
LoCoMo	Dialogue turn	0.72
	Observation	0.95
	Event summary	0.88

Appendix

Table 7: Source reliability weights used for evidence aggregation. The weights are fixed heuristic values determined by evidence provenance and granularity. They are not calibrated probabilities of factual correctness and are not tuned on evaluation instances.

Weighting Strategy	Accuracy (%)
No weighting ( $\nu=1$ )	78.4
Uniform weighting ( $\nu=0.75$ )	78.9
Source-specific weighting	79.6

Appendix

Table 8: Sensitivity analysis of source reliability weighting on LongMemEval-S with Qwen3-30B-A3B-Instruct-2507.

Relation	Evidence considered	Operational scoring rule	Decision criterion
Support	Exact matches after lowercasing and whitespace normalization; accumulated evidence and source confidence.	Identical normalized records are merged, with evidence counts and confidence updated. No independent support-edge score $f_{\mathrm{sup}}$ is computed.	Exact normalized match; no $\theta_{\mathrm{sup}}$ .
Conflict	Matching entity–attribute slots or compatible personal claims; preference polarity, constraints, location changes, and task status.	$f_{\mathrm{conf}}=0.95$ (preference reversal), $0.94$ (temporal location change), $0.92$ (constraint change), $0.90$ (task-status change), $0.88$ (slot-value update), or $0.70$ (weak opposing signals with overlap $o\geq 0.45$ ).	All detected conflicts are linked. Strong conflict: $f_{\mathrm{conf}}\geq 0.75$ and non-weak type.
Dependency	Explicit source-memory IDs, shared structured slots, dialogue-turn provenance, predefined semantic cues, and token overlap.	$f_{\mathrm{dep}}=0.95$ (explicit reference); $0.55$ (same slot, non-derived); $\max(0.90,o)$ (derived, same turn, $o\geq 0.12$ ); $\max(0.80,o,s)$ (derived, semantic cue $s\geq 0.75$ ); or $0.70$ (derived, topic overlap $o\geq 0.45$ ).	Link if a rule matches; no global $\theta_{\mathrm{dep}}$ . Verify at cascade risk $D_{i}\geq 0.70$ ; mark stale for superseded sources with $f_{\mathrm{dep}}\geq 0.80$ .

Appendix

Table 9: Operational relation construction and decision criteria in the CALM implementation.

Dataset	Category	Questions
LongMemEval-S	SS-User	70
	SS-Pref.	30
	SS-Asst.	56
	K-Upd.	78
	Multi	133
	Temp.	133

Appendix

Table 10: Question distributions of the evaluated benchmark subsets. LoCoMo adversarial questions are excluded from the reported results.

Stage	Hit $\uparrow$	FullHit $\uparrow$	AvgRecall $\uparrow$
Initial Retrieval	100.00	94.80	97.99
Post-Gating	100.00	94.60	97.84
Final Packing	100.00	93.00	97.18
Additional Metric			Value
Packed-Hit Accuracy			79.6

Appendix

Table 11: Evidence-coverage diagnostics across retrieval, gating, and context packing. All values are reported as percentages.

Post-retrieval policy	Acc. (%)
Validity-Rerank	77.20
CALM-ZS	76.60
CALM	79.60
Hard Filter + Relevance	76.00
Hard Filter + Validity-Rerank	78.00
Hard Filter + Learned Rerank	78.40

Appendix

Table 12: Illustrative comparison of matched post-retrieval admission controls on LongMemEval-S. All methods share retrieved candidates, answer model, and memory-token budget.

Variant	Erroneous-Memory Injection Ratio
	0%	10%	20%	30%	40%	50%
CALM	79.8	78.8	77.6	76.2	74.6	72.8
Relevance-Only	75.4	74.6	72.2	69.6	66.8	56.6
w/o Stage III	74.0	71.2	68.0	64.6	61.0	57.2

Appendix

Table 13: Accuracy (%) under increasing erroneous-memory injection ratios. The candidate-set size and final memory-token budget are held constant across methods. Bold values indicate the best performance at each injection level.

Method	512	1024	2048	Full
EverMemOS	29.60	53.80	62.40	66.40
CALM-ZS	38.00	58.40	69.80	76.60
CALM	39.20	60.20	74.60	79.60

Appendix

Table 14: Exact accuracy (%) under different memory-token budgets on LongMemEval-S.

Persistent Memory for Language Models

Persistent memory for language models studies how models retain, update, and use information across interactions and beyond their context windows.

See all 272 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.
RaMem: Contextual Reinstatement for Long-term Agentic Memory
Jun 22, 2026Wei Yang, Bryce Kan, Shixuan Li +5Agent Memory LLM Agent Memory
University of Southern California
EMIR $^2$ : Evolution-Aware Memory with Intent-Guided Multi-Round Retrieval
Sep 26, 2026Jinlan Liu, Hongliang Sun, Yong Wang +4Memory Management for LLM Agents LLM Agent Memory
Department of Computer Science, Weihai & Qingdao Research Institute Harbin Institute of Technology Weihai, China

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

Persistent Memory for Language Models

Explore similar work

MemoryLACE: Memory Lifecycle-Aware Consolidation and Evidence Retrieval

RaMem: Contextual Reinstatement for Long-term Agentic Memory

EMIR2^22: Evolution-Aware Memory with Intent-Guided Multi-Round Retrieval

EMIR $^2$ : Evolution-Aware Memory with Intent-Guided Multi-Round Retrieval