← sources

Preprint: in two agent memory setups, revoked policies ranked above their replacements

source: reddit (reddit:AI_Agents)published 2026-11-02 03:18ingested 2026-11-03 02:37

retained text

This is the authors' claim from a preprint; I haven't tested anything.

Some agent memory systems mark a contradicted fact as invalid and keep it instead of deleting it. The paper tested five (Graphiti, Zep, mem0, langmem, cognee) by loading each with an old policy that permits a harmful action plus the newer policy that forbids it, across nine scenarios and nine models.

What they report: no system enforced revocation in every case. Only two setups returned the revoked record: Graphiti, and mem0 with its read-time filter turned off by the authors (mem0 filters revoked records by default). In both, the revoked record ranked first in every scenario, and agents took the unsafe action in more than two of every five trials. cognee and langmem, in the direct-insertion test, deleted the old record outright.

The authors also built a guard between the agent and the memory backend that withholds revoked or contradicted records.

If you run agent memory: do you delete outdated facts, or keep them flagged and rely on the flag being checked at retrieval?

document fetched 2026-11-03 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xlmhkc/preprint_in_two_agent_memory_setups_revoked