MemHarness: Memory Is Reconstructed, Not Replayed

MemHarness is a memory-augmented framework that lets LLM agents reconstruct retrieved experiences to fit the current context instead of replaying them verbatim.

A policy model critiques and adapts recalled experiences before acting, making memory use more context-aware.

Trained end-to-end with GRPO, it was evaluated on ALFWorld and WebShop and outperformed static-memory baselines and pure RL.

The approach also improved robustness in out-of-distribution settings, reducing negative transfer from rigid memory replay.