MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents
Boyu Yang, Jiazheng Sun, Zilong Lu, Zhi Qiu, Xin Peng, Jun Zheng
cs.AI · cs.SE
Quick verdict
Shapley-based memory clearance hits 82.3% recovery on injected LoCoMo faults; theory sound but gains over ContextCite are within noise, cost unreported.
Summary
MeClear is a post-retrieval memory-clearance framework for long-horizon LLM agents. Given a frozen retrieved context, it screens records with Leave-One-Out counterfactuals, estimates each memory's contribution via permutation-sampled Shapley attribution over a binary task-value function, ranks suspected harmful records, and clears a minimal nested candidate set, verifying task recovery before answering while leaving the persistent memory bank untouched. Theorem 1 formalizes why LOO misses redundant harm (zero LOO effect but -Δ/2 Shapley contribution), Proposition 1 gives unbiasedness and Hoeffding concentration for the sampled estimator, and Theorem 2 guarantees empirical non-degradation within the nested family. Experiments use LoCoMo: 745 behaviorally verified cases with 1,115 injected faults (375 direct, 285 redundant, 85 joint) over 10 conversations, Mem0 2.0.12+Qdrant memory, Kimi-k2.6 agent, Qwen3.6-Flash judge. MeClear reports 85.9% target recall, 47.0% exact set match, and 82.3% binary recovery versus LOO's 38.3%/34.2%/56.8% (+25.5pp recovery). However, against ContextCite the recovery edge is +0.8pp with CIs crossing zero (win/loss 66/60), and ProxySPEX wins exact match 54.2% vs 47.0%. Main limitations: faults are synthetic LLM-injected paraphrase conflicts filtered to 14.8% causal validity; only LoCoMo; no cost accounting despite roughly a hundred agent calls per query at K=5, L=16; the recovery metric is judged by the same evaluator the method verifies against; and no ablation isolates the verification wrapper's contribution from attribution quality.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Sampled cooperative Shapley attribution identifies redundant and jointly harmful memories that Leave-One-Out deletion provably misses (Theorem 1: zero LOO effect vs -Δ/2 Shapley contribution).
- Query-scoped verified clearance guarantees empirical non-degradation and verified recovery within the nested candidate family, without modifying the persistent memory bank (Theorem 2).
- On 745 causally verified LoCoMo cases with 1,115 injected faults, MeClear achieves 85.9% target recall, 47.0% exact set match, and 82.3% binary task recovery, +25.5pp over LOO.
- Performance gaps concentrate in multi-memory fault regimes: redundant-conflict recovery rises from 6.3% (LOO) to 68.4%, and joint-interaction recovery reaches 91.8%.
Strengths
Clean formalization of the LOO blind spot
Theorem 1 gives a provable account of redundant-harm masking (zero LOO effect, Shapley contribution -Δ/2), extended to r-member groups in Corollary A.1; the plateau/floor signatures in Table 1 match the theory's predictions exactly.
Above-average evaluation hygiene
745 paired cases pass per-fault behavioral admission criteria (Table 3 funnel: 7,986 proposals → 1,178 causal-valid → 745); 95% CIs use 2,000 two-level cluster-bootstrap replicates over conversations and queries; n=5 cells are flagged descriptive and 2 exclusions are disclosed.
Safety-conscious intervention design
Gains concentrate exactly where the mechanism predicts
Budget sensitivity supports the mechanism, not over-deletion
Weaknesses
Headline built on the weakest baseline
Verification wrapper may drive the recovery metric
Synthetic, behaviorally pre-filtered faults
No cost accounting
LOO screening is decorative in the decision path
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MeClear (arXiv 2609.09115; Fudan SE Lab + BIT) targets a real failure mode of memory-augmented agents: retrieval by semantic similarity surfaces stale, redundant, or mutually reinforcing harmful records whose damage only manifests jointly. Claimed contributions are (1) a task-conditioned clearance framework combining LOO screening with permutation-sampled Shapley attribution, (2) three theoretical results — Theorem 1 (redundant-harm blind spot), Proposition 1 (finite-sample unbiasedness + Hoeffding concentration), Theorem 2 (verified non-degradation), and (3) experiments on 745 behaviorally verified LoCoMo cases reporting 85.9% target recall and 82.3% binary task recovery, +25.5pp over LOO. The 'first to apply cooperative game-theoretic allocation to harmful memory clearance' framing sits awkwardly beside the paper's own citations of Shapley-based RAG source attribution (Nematov et al. 2025) and its ProxySPEX baseline.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MeClear?
MeClear is a post-retrieval memory-clearance method for long-horizon LLM agents. It estimates each retrieved memory's contribution to task success using permutation-sampled Shapley cooperative-game attribution, then temporarily masks the most harmful records for the current query, verifying that the agent's answers recover before responding — all without permanently editing the memory store.
How much does MeClear improve over baselines?
On 745 injected-fault cases built from LoCoMo dialogues, it reports 85.9% harmful-memory recall and 82.3% binary task recovery — 25.5 percentage points above the Leave-One-Out baseline. Against the stronger ContextCite baseline, however, the recovery edge is only 0.8 points with confidence intervals crossing zero, and ProxySPEX achieves higher exact-set-match precision (54.2% vs 47.0%).
Does MeClear delete memories permanently?
No. Clearance is query-scoped and reversible: identified records are suppressed only from the active context for the current query after a recovery check, and the persistent memory bank is left intact by design to avoid cross-task regression.
What data and models were used to evaluate MeClear?
Ten LoCoMo conversations (368 clean queries, 745 verified cases) with 1,115 LLM-injected and behaviorally verified fault records. Kimi-k2.6 is the task agent, Qwen3.6-Flash the judge, Mem0 2.0.12 with Qdrant the memory system (top-5 retrieval), and comparison baselines are Leave-One-Out, ContextCite, ProxySPEX, and direct LLM filtering.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Numeric labels in Figures 4, 5, 7, 10-12 are interleaved blocks in the PDF extraction; values were mapped to methods by matching the narrative text (which repeats the key figures verbatim) — minor misassignment risk applies only to rarely-cited cells.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.