MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents
MeClear tries to keep long-running agents from using memories that hurt the current task.
The paper frames the problem as retrieval systems pulling in memories that are semantically related but outdated, misleading, or conflicting. MeClear scores memory utility with Leave-One-Out screening and sampled Shapley attribution, then suppresses harmful evidence for the query without changing the stored memory bank. Across ten long-dialogue memory pools, the authors report 85.9% target recall and 82.3% task recovery, 25.5 points above their LOO baseline. ArXiv · AI/CL/LG's note
The paper frames the problem as retrieval systems pulling in memories that are semantically related but outdated, misleading, or conflicting. MeClear scores memory utility with Leave-One-Out screening and sampled Shapley attribution, then suppresses harmful evidence for the query without changing the stored memory bank. Across ten long-dialogue memory pools, the authors report 85.9% target recall and 82.3% task recovery, 25.5 points above their LOO baseline. ArXiv · AI/CL/LG's note
score 4