An agent memory store is a persistent layer that saves an AI agent’s learned state, preferences, workflow context, or past observations so they can be reused later. Unlike a short-lived prompt, memory survives across sessions and can help an agent act more consistently, remember approved tools, or continue a task after interruption.
In cyber security, this persistence is both useful and risky. If an attacker can read or write the memory store, they may steal secrets, alter instructions, or poison the agent’s behavior for future runs. That can turn a single compromise into long-lived persistence. Defenders should treat memory as sensitive data: restrict access, encrypt it, validate what can be written, and monitor for unexpected changes. In AI orchestration systems, memory stores often sit close to tool access and credentials, so compromise of the surrounding bridge can expose more than just model output.