Persistent agent memory is data an AI agent saves across sessions so it can reuse context later. That memory may include preferences, task state, identifiers, tool history, or other notes that help the agent behave consistently over time. In security terms, it turns an AI system from a stateless chat into a stateful control surface.
This matters because saved memory can become a target for theft, tampering, or leakage. An attacker who reaches the memory store may learn secrets, influence future outputs, or plant malicious instructions that survive a restart. Defenders should treat agent memory like sensitive application state: limit what is stored, separate it from credentials, authenticate access to the memory service, log reads and writes, and validate or sanitize any remembered content before it is reused.