Suppressing memories with negative downstream utility beat leave-one-out screening by 25.5 points on task recovery

arXiv 2609.09115·medium signalXBlueskyLinkedInCopy link

MeClear targets a failure mode in retrieval-based agent memory: retrievers optimize semantic similarity, not usefulness, so outdated, misleading and conflicting evidence gets pulled into the active context. It combines leave-one-out screening with sampled cooperative Shapley attribution to spread utility across interacting evidence, which resolves the redundant-conflict masking that defeats single-removal evaluation, then applies a query-scoped minimal clearance over a nested filtration and verifies task recovery on the cleared context without mutating the persistent memory bank. Across ten long-dialogue memory pools it reached 85.9% target recall and 82.3% task recovery, 25.5 points above leave-one-out baselines.