Definition
An attack where a malicious actor plants false or harmful information into the memory that an AI agent retains across conversations. Because the agent automatically recalls and acts on this stored memory in future sessions, a single poisoning event can permanently alter the agent's behaviour — silently redirecting its actions, leaking data, or undermining trust long after the attacker has gone.
Why it matters
Enterprise AI agents increasingly remember context across sessions to be more helpful; this feature becomes a persistent back door if not secured. A single compromised memory entry can corrupt every future interaction that agent has with every user.