What they did

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. This is intended to detect cases where redundant harmful memories mask one another because removing either item alone does not improve performance.

MeClear ranks memories by their attributed utility and tests progressively filtered contexts to find a minimal, query-specific clearance that restores task performance. The suppression applies only to the current query and does not delete items from the persistent memory bank; experiments covered ten long-dialogue memory pools.

Key findings

• MeClear achieved 85.9% target recall for the memories it sought to identify.
• The method restored successful task execution in 82.3% of evaluated cases.
• Task recovery was 25.5 percentage points higher than with leave-one-out clearance alone.
• Cooperative attribution addressed interacting or redundant conflicts that single-removal tests can miss.

Why it matters

Long-horizon agents can fail not because relevant information is absent, but because retrieved memory contains harmful evidence. MeClear reframes memory management as task-conditioned utility estimation rather than semantic matching alone, offering a way to correct the active context without permanently erasing stored information.

Caveats

The reported evaluation spans ten dialogue memory pools, so its generality across models, memory architectures, domains, and substantially larger stores remains uncertain. Sampled Shapley estimation and repeated task verification may also add computational cost, while the abstract does not report latency, token usage, or sensitivity to the underlying utility evaluator.