Shanghai AI Lab Open-Sources MemHarness: Critical Memory for Agents

aigclink · x · 2026-08-05

Shanghai AI Lab has open-sourced MemHarness, an AI memory enhancement framework designed to solve the 'negative transfer' problem where agents blindly apply outdated experiences to new scenarios.

Its core mechanism forcefully inserts Critic and Reconstruct modules between memory retrieval and action execution. It actively compares historical states with current states, discarding irrelevant memories if the discrepancy is too large, enabling human-like 'critical reference'.

The framework requires zero manual annotation, relying entirely on the GRPO algorithm combined with task rewards for end-to-end training, which significantly reduces the cost of migrating to new tasks or domains.

Related event: Shanghai AI Lab Open-Sources MemHarness Framework(2 posts)→

Original post →

More from coding & agent

coding & agent channel →