Shanghai AI Lab Open-Sources MemHarness for Critical Agent Memory

aigclink · x · 2026-08-05

Shanghai AI Lab has open-sourced MemHarness, a memory-augmented framework that empowers LLM agents to actively reconstruct past experiences based on current context rather than verbatim replay.

Traditional memory-augmented agents treat retrieved memories as static records, which can cause negative transfer when past experiences misalign with the current state. To solve this, MemHarness introduces Critic and Reconstruct modules between retrieval and action. These modules actively compare past and present states, discarding experiences if the discrepancy is too large.

The framework is trained end-to-end via GRPO with task rewards, requiring no manual annotation. This significantly lowers the migration cost when adapting the agent 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 →