MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance
Boyu Yang, Jiazheng Sun, Zilong Lu, Zhi Qiu, Xin Peng, Jun Zheng — Fudan University · Beijing Institute of Technology
Key Insight
Reports 85.9% target recall and 82.3% task recovery across long-dialogue pools when clearing harmful/redundant agent memory.
What's Actually Supported
Shapley-style attribution can identify which memories to clear with high target recall, 25.5 points above a leave-one-out baseline, per the paper's own long-dialogue evaluations.
Caveat / Limitations
Game-theoretic estimation adds real compute cost; evaluations use curated dialogue pools and don't prove durable privacy or revocation guarantees.
