上海AI Lab团队推出MemHarness：让Agent记忆像人类一样被重构

TL;DR - MemHarness is an LLM-agent framework that evaluates and rewrites retrieved experiences for the current context before acting. This reduces harmful transfer from stale memories and improves decision-making, particularly in out-of-distribution settings.

• Adds explicit retrieval, critique/reconstruction, and action-generation stages instead of injecting memories unchanged.
• Trains the full workflow end-to-end with GRPO using task rewards and format constraints, without reconstruction labels.
• Outperforms pure-RL and static-memory baselines on ALFWorld and WebShop; removing reconstruction erases much of the gain.
• Experiments indicate that source/current-state compatibility determines whether memories should be retained, rewritten, or discarded.