Current memory tools can pass a poisoned run on. Procedural memory can stop it (github.com)

A recent study examined the efficacy of various memory configurations in AI models, particularly focusing on whether skills learned from a poisoned session can negatively impact future learning sessions. Researchers analyzed three models—Claude Code, Claude Haiku 5.5, and Claude Sonnet 5.5—by employing four memory strategies: no memory, naive memory, replay, and "MuscleMemory." Their findings revealed that while naive memory retained and trusted all learned skills, it also allowed the propagation of poisoned inputs across sessions. In contrast, the MuscleMemory approach effectively mitigated the poison, resulting in zero instances of the poison being served in tested sessions, highlighting its robustness. This research is significant for the AI/ML community as it deepens the understanding of memory systems' vulnerabilities, particularly in the context of adversarial inputs or "poisons." The study underscores the importance of developing memory architectures that can distinguish between harmful and beneficial inputs. Notably, while the naive memory strategy can appear effective at first glance, it ultimately fails to prevent the distribution of poisoned knowledge, which emphasizes the critical need for innovative memory models that safeguard against such risks in AI learning environments.