MemHarness: Reconstructing Memory Instead of Replaying Makes Agents More Reliable

rohanpaul_ai · x · 2026-08-05

Traditional memory-augmented agents often paste semantically similar past experiences directly into the context, which can fail or become noise if the state has changed.

MemHarness introduces a memory reconstruction mechanism, adding a decision step between retrieval and action: the agent compares the memory's original state with the current one, rewrites the transferable parts, or rejects it entirely to reason without memory. This reconstruction is learned end-to-end via Group Relative Policy Optimization (GRPO) based on task rewards.

Experiments show that this method significantly outperforms the memory-free RL baseline on ALFWorld and WebShop. Interestingly, even when memory is disabled at test time, the policy trained with this method still maintains a high success rate, suggesting that memory reconstruction not only improves inference-time adaptability but also teaches the policy itself to judge situations more carefully.

Original post →

More from coding & agent

coding & agent channel →