Memory poisoning is the corruption of an AI agent's persistent memory with false, manipulated or attacker-supplied content, so that the agent's later behaviour is altered by material it now treats as its own established knowledge.
What is memory poisoning?
Agent memory is unusually trusted: content that has been written to memory is typically recalled without the scrutiny applied to freshly retrieved documents. That makes it a durable attack surface — a single successful injection can persist across sessions and users, and the effect appears long after the event that caused it. Defending memory means governing writes as strictly as reads, retaining the provenance of every entry, scoping recall by identity and purpose, and being able to expire or revoke entries once they are found to be wrong.