MemHarness: Memory Is Reconstructed, Not Replayed
MemHarness trains agents to adapt retrieved memories to the current state before acting.
The paper argues that replaying past experiences verbatim can hurt LLM agents when the old situation does not match the new one. MemHarness uses a unified policy model to critique and reconstruct retrieved experience into context-grounded guidance at each decision step. The authors report gains over pure RL and static memory baselines on ALFWorld and WebShop, including out-of-distribution tests. HF Daily Papers' note
The paper argues that replaying past experiences verbatim can hurt LLM agents when the old situation does not match the new one. MemHarness uses a unified policy model to critique and reconstruct retrieved experience into context-grounded guidance at each decision step. The authors report gains over pure RL and static memory baselines on ALFWorld and WebShop, including out-of-distribution tests. HF Daily Papers' note
score 4