MemHarness: Memory Is Reconstructed, Not Replayed

TL;DR - MemHarness trains LLM agents to reconstruct retrieved experiences for their current state rather than replaying static memories verbatim. This reduces negative transfer and improves performance and out-of-distribution robustness.

• A unified policy critiques and adapts retrieved experiences into context-grounded guidance at each decision step.
• The reconstructive behavior emerges through end-to-end GRPO training.
• On ALFWorld and WebShop, MemHarness outperforms pure reinforcement learning and static memory-augmented baselines.
• Analyses suggest reconstruction also provides latent training guidance that improves intrinsic reasoning.