AI agents that save memories from past conversations might be misremembering them, and a new audit tool sets out to prove it.
The paper, "Memory Is a Derivation: The Distributed-Evidence Paradox in Long-Term Agents," was posted September 30, 2026 to arXiv's cs.AI category as arXiv:2609.36130v1. The researchers built DerivAudit, a framework that checks whether a long-running agent's saved memory is actually backed by the conversation history available at the moment it was written. Testing on two memory corpora, they found that citations attached to a memory often understate its real support: pulling in broader pre-write history recovered justification for nearly 60% of memories that looked unsupported at first glance. But 17-21% stayed unsupported even after that expansion, and simply adding more context did not fix the underlying problem - unsupported memories still got admitted across multiple verification models, and on two of the backbones tested, broader evidence actually made admission worse.
Persistent memory is the feature vendors pitch as the fix for agents that forget everything between sessions. This paper shows memory can also compound errors, since an agent may combine individually true facts into a composite claim the original conversation never established. That is a subtler failure than a chatbot inventing a fact outright - it is an agent building a citation trail for something it never actually proved.
A longer memory does not just mean an agent remembers more. It means the agent gets more chances to quietly reason its way to a conclusion nobody checked.