Shanghai AI Lab team launches MemHarness: Reconstructing agent memory like humans - 36Kr

Shanghai AI Lab and collaborators introduced MemHarness, which makes LLM agents critique and reconstruct retrieved memories before acting.

With end-to-end GRPO training, agents can rewrite, keep, or drop past experiences based on the current context instead of pasting them into prompts.

MemHarness outperformed pure reinforcement learning and static memory-augmented baselines on ALFWorld and WebShop.

It was also more robust in out-of-distribution settings, reducing negative transfer and improving generalization and reasoning.