Presents MeClear, a task-conditioned memory clearance method that removes memories with negative downstream utility via cooperative attribution, improving long-horizon agent reliability.
MeClear is a framework for long-horizon LLM agents that improves agent reliability by identifying and suppressing memories with negative downstream utility. It utilizes cooperative game-theoretic attribution (specifically sampled Shapley values) to resolve "redundant conflict masking" where traditional single-removal methods fail to detect harmful evidence.
Key performance metrics include: 85.9% target recall for memory identification. 82.3% overall task recovery rate. * A 25.5 percentage point improvement over Leave-One-Out (LOO) baselines.
The method operates as a non-destructive, query-scoped clearance layer that freezes the active context, evaluates memory contributions, and selectively suppresses harmful records without permanently altering the persistent memory bank.
MeClear targets a core failure mode in long-horizon LLM agents: persistent memory can improve continuity, but it also accumulates stale, conflicting, or misleading entries that degrade later decisions. The work frames memory management not as simple eviction by recency, importance, or storage budget, but as a task-conditioned clearance problem in which the value of a memory depends on how it affects downstream task performance. The central mechanism is cooperative game-theoretic attribution, which estimates each memory’s contribution to the agent’s utility by accounting for interactions among memories—such as redundancy, complementarity, or conflict—rather than scoring entries in isolation.
The key contribution is a principled bridge between credit assignment and memory lifecycle control. By treating memories as contributors to task outcomes, MeClear can identify entries with negative expected utility and clear them in a risk-aware manner, i.e., only when the anticipated benefit of removal outweighs the uncertainty or potential harm to future performance. This shifts memory management from heuristic pruning to an optimization-oriented process: the agent maintains a compact, task-aligned memory state while avoiding the deletion of entries that may be useful under other conditions.
This matters because long-horizon agents increasingly depend on memory to operate across sessions, tasks, and extended decision horizons. Uncontrolled memory growth introduces retrieval noise, policy drift, privacy risk, and compounding errors. MeClear is relevant because it offers a more systematic way to keep agent memory reliable and task-relevant, improving robustness in settings such as multi-session assistants, autonomous workflows, and agents that must make decisions over days or weeks rather than single episodes.