Most memory-augmented AI agents treat retrieved experiences as static records, replaying them verbatim into the context. This approach often fails because stored memories are abstract and general, while the agent's current situation is concrete and ever-changing. The mismatch can cause negative transfer, where irrelevant memories hurt performance.
Inspired by human memory, which reorganizes past experiences using present cues rather than reproducing them exactly, researchers propose MemHarness. This framework inserts explicit memory critique and reconstruction between retrieval and action, turning static records into context-sensitive guidance while keeping traceability.
MemHarness operates in five stages: observation, retrieval, critique, reconstruction, and action. A unified policy model critiques and reconstructs the retrieved experience based on the current state, producing grounded guidance before acting. This reconstructive ability emerges naturally through end-to-end training with GRPO, requiring no human annotation.
Experiments on ALFWorld and WebShop show that MemHarness substantially outperforms both pure RL and static memory-augmented baselines, with strong robustness in out-of-distribution scenarios. Ablations confirm that adaptive reconstruction—not retrieval alone—is the primary driver of the gains. The reconstruction objective also serves as latent guidance during training, improving the agent's intrinsic reasoning capabilities.