MeClear: Cooperative Game-Theoretic Attribution and... | AI Research

MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents
Long-horizon Large Language Model (LLM) agents rely on external memory to store information across extended interactions. However, current systems often retrieve memories based solely on semantic similarity, which can lead to the inclusion of outdated, conflicting, or misleading information that degrades the agent's performance. MeClear is a framework designed to identify and suppress these harmful memories by evaluating their actual impact on task performance. Instead of permanently altering the memory database, MeClear dynamically filters the active context for each specific query, ensuring that only useful information influences the agent's output. The ai search story also surfaces in Stanford AI discovery identifies natural weight..., adding another angle.

Identifying Harmful Memories

The core challenge in memory management is that memories often interact in complex ways. A single piece of information might be harmless on its own but become problematic when combined with other records, or two redundant memories might mask each other's negative effects. MeClear addresses this by treating the memory context as a cooperative game. It uses a combination of "Leave One Out" (LOO) screening—which tests the effect of removing one record at a time—and sampled cooperative Shapley attribution. This allows the system to measure the marginal contribution of each memory across various combinations, effectively uncovering "hidden" toxicity that simple deletion methods miss.

Verified Context Clearance

Once MeClear identifies which memories are negatively impacting a task, it performs a query-scoped clearance. It follows a "minimal intervention" principle: the system seeks to maximize task recovery while removing the smallest number of records necessary. Crucially, this process is verified before the agent generates its final response. By testing the cleared context, MeClear ensures that the removal actually improves performance. Because this clearance is applied only to the active context during inference, the original, persistent memory bank remains intact, preventing the risk of permanent data loss or cross-task performance regression. The same ai search question is explored in Measure Before You Manage, which adds a research perspective.

Performance and Results

MeClear was evaluated across ten different long-dialogue memory pools to test its effectiveness in maintaining reliable agent behavior. The results demonstrate that the framework is highly capable of identifying and mitigating harmful memory influences. MeClear achieved a target recall of 85.9% and an overall task recovery rate of 82.3%. These figures represent a 25.5 percentage point improvement over standard Leave One Out baselines, highlighting the effectiveness of using cooperative game theory to manage the complex dependencies found in long-term agent memory. The ai agents story also surfaces in OpenAI Unveils GPT-Red an Automated Model..., adding another angle. as detailed in the full paper on Arxiv