arXivPaperNeeds Review

MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents

MeClear suppresses memories that hurt the current task, using cooperative attribution to expose conflicts that single-memory removal can miss, while leaving the persistent memory bank unchanged.

arXiv · Sep 8, 2026

Open Source Open Markdown Open JSON

Source Summary

MeClear scores memories by downstream utility using Leave One Out screening plus sampled Shapley attribution, then tests progressively filtered contexts. Across **10 long-dialogue memory pools**, it reached **85.9% target recall** and **82.3% task recovery**.

Practical Implication

Builders of long-running agents should treat retrieval relevance as insufficient: retrieved memories can be stale, misleading, or mutually conflicting. Query-scoped suppression offers a pattern for recovering a task without deleting durable memory.

Agent-Ready Context

MeClear scores memories by downstream utility using Leave One Out screening plus sampled Shapley attribution, then tests progressively filtered contexts. Across **10 long-dialogue memory pools**, it reached **85.9% target recall** and **82.3% task recovery**.

Builders of long-running agents should treat retrieval relevance as insufficient: retrieved memories can be stale, misleading, or mutually conflicting. Query-scoped suppression offers a pattern for recovering a task without deleting durable memory.

The reported recovery rate is a **25.5 percentage-point improvement over LOO baselines**, but the supplied material covers only the authors’ experiments. Runtime cost, sensitivity to the utility test, and performance on production memory stores remain open.

Connected Context · Feed7 Judgment

MeClear shifts memory reliability from deciding what to retrieve or retain to testing whether retrieved combinations actually help the current task. Its query-scoped suppression provides a reversible response to stale, misleading, or conflicting context, strengthening prior warnings that relevant or reusable memories can still degrade performance. The reported gains support utility-aware filtering in controlled long dialogues, while production cost and portability remain unresolved.

Break It Down, Pass It On: Cross-Task Skill Transfer in LLM AgentsBoth show that retrieved memory can perform worse than having no memory; MeClear adds a query-time attribution and suppression mechanism for identifying and withholding harmful memories.PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM AgentsPsychoAgent reranks memories for conflict-sensitive salience, whereas MeClear evaluates their downstream task utility and progressively removes harmful context, offering a complementary control after relevance ranking.Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory PortabilityMemory portability identifies model upgrades as a source of retrieval failure; MeClear could mitigate harmful migrated context at query time, but does not replace embedding rebuilds or migration testing.thedotmack/claude-memClaude-Mem supplies an inspectable persistent retrieval workflow but leaves contradiction handling unresolved; MeClear provides a compatible pattern for temporarily suppressing damaging retrieved observations without deleting durable history.

Context Map

agentcoding#agent-memory#retrieval#agent-reliability

Uncertainty

The reported recovery rate is a **25.5 percentage-point improvement over LOO baselines**, but the supplied material covers only the authors’ experiments. Runtime cost, sensitivity to the utility test, and performance on production memory stores remain open.