Persistent memory lets one successful injection keep influencing future sessions. Know how memory is written, give users and admins a way to review and clear it, and watch for memory writes triggered by untrusted content.
Read in this order
- OfficialAgentic AI: Threats and Mitigations (OWASP GenAI Security Project)Start here: where memory poisoning sits among agent threats and mitigations.
- ArticleChatGPT: Hacking Memories with Prompt Injection (Embrace The Red)A concrete exploit that planted false memories through prompt injection.
- PaperAgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge BasesResearch showing poisoned memory entries can steer agent behavior.