For LLM agents with long-term memory, removing harmful context is more important than retrieving similar context—MeClear uses cooperative game theory to identify which memories hurt performance and filters them on-the-fly.
MeClear is a framework that identifies and removes unhelpful memories from LLM agents' external memory systems during long conversations. Instead of just retrieving semantically similar memories, it uses game-theoretic methods to detect which stored information actually hurts task performance, then temporarily filters them out without permanently deleting anything.