MeClear is a memory clearance framework for long-horizon LLM agents. It decides which retrieved memories to keep in active context by estimating how much each memory contributes to task success, using Leave One Out screening and sampled cooperative Shapley attribution. This catches cases where several memories jointly cause failure even though removing any single one makes no difference. On ten long dialogue memory pools, MeClear reached 85.9% target recall and 82.3% overall task recovery, a 25.5 percentage point improvement over Leave One Out baselines. It suppresses harmful memories temporarily at query time, verifies that the cleared context restores task behavior, and leaves the persistent memory bank unchanged.
Game-theoretic filtering helps agents suppress harmful memories during long tasks
MeClear attributes downstream utility across interacting memories, then temporarily removes the smallest harmful subset needed to restore task performance.
Independent
Boyu Yang · Jiazheng Sun · Zilong Lu · Zhi Qiu · Xin Peng · Jun Zheng
Research Digest··2 min read
Thread:Memory Management for Agents
In one line
Long-horizon LLM agents can selectively suppress harmful retrieved memories via cooperative Shapley attribution, achieving 85.9% target recall and 82.3% task recovery.
Before you read this
Three ideas this paper builds onagent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.200 papers →credit assignmentDetermining which actions or steps in a sequence contributed to a final outcome, often for reinforcement learning.171 papers →failure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.115 papers →
Read it as
Yang et al. developed MeClear to identify retrieved memories that impair an LLM agent, including outdated, misleading, or mutually conflicting evidence.
The authors combined leave-one-out screening with sampled Shapley attribution, a cooperative game-theoretic method that assigns utility while accounting for interactions among memories.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.