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

What it claims

CALM is a memory layer that sits between an existing memory backend's retriever and the answer model. It decides which retrieved memories enter the working context. It does not change the retriever or the LLM. The authors, from Harbin Institute of Technology and Meituan (arXiv, 28 October 2026), separate three properties: semantic relevance, query-independent validity, and query-conditioned eligibility. Superseded memories are kept in storage and blocked only when a query asks about the current state. The best reported figure is 79.60% on LongMemEval-S with Qwen3-30B-A3B-Instruct-2507, described as best or tied-best overall accuracy.

How it works

Stage I converts backend records (dialogue fragments, extracted facts, summaries) into canonical cells. Each cell holds content, evidence, semantic type, temporal anchor, provenance, entities, applicability scope, a reliability weight and a lifecycle state. New cells start as CANDIDATE.

Stage II runs deterministic rules, not a learned scorer:

• Identical normalized claims are merged and their evidence counts updated. There is no separate support-edge score.
• A conflict with strength of at least 0.75 can trigger supersession. Weaker conflicts are recorded without replacing either claim.
• Direct user evidence or confidence of at least 0.60 makes a cell active. Repeated support or confidence of at least 0.80 makes it STABLE.
• Strong supersession caps the old memory's stored validity at 0.15.
• Elapsed time alone never marks a memory STALE. Time only orders updates and sets applicability intervals.
• An invalid dependency can trigger verification or mark the dependent memory STALE. Archival is an optional retention policy.

Stage III scores each retrieved memory on four features: relevance, stored validity, query-memory compatibility, and 1 minus query-conditioned risk. The risk term covers conflict, dependency, temporal and lifecycle risk.

flowchart LR
  R[Retrieved candidates] --> H{Hard eligibility filter}
  H -- fail --> K[Blocked, kept in storage]
  H -- pass --> Z[CALM-ZS: mean of 4 features]
  H -- pass --> L[CALM: MLP gate, threshold 0.50]
  Z --> T{Score above threshold?}
  L --> T
  T -- no --> K
  T -- yes --> C[Ranked, packed into token budget]

CALM-ZS is training-free and weights all four features at 1/4. The learned variant is an MLP trained only on development instances with separately built query-memory labels, and its threshold is 0.50. Admitted memories are ordered by score and packed under a fixed memory-token budget.

Evidence

Evaluation uses LongMemEval-S (500 questions, six categories) and LoCoMo (1,986 questions over ten conversations), with Qwen and GPT answer models. Three backbones are reported in total. The comparison baselines include HippoRAG, A-Mem, Mem0, Zep and A-TMA. The page also reports:

• Simple alternatives: query-conditioned admission improves evidence selection over plain filtering and reranking.
• Conflict blocking and robustness: erroneous retrieved memories are suppressed while useful ones are kept.
• Historical reconstruction: earlier-state questions are answered correctly without hurting current-state answers.
• Other analyses: component ablations and efficiency results.

Limits

The fetched text stops at the start of the experimental setup. Baseline scores, per-category results, the other backbones' numbers, the CALM-ZS threshold, the hard-filter definition and the ablation and efficiency figures do not appear. Those details are deferred to Appendix A.1 to A.4 and the results tables. Conflict and dependency detection are rule-based, so coverage depends on rules that are not shown. The page mentions no code release or licence.