AI Agent Memory Poisoning Defense Checklist

A poisoned instruction can wait weeks before it acts.

Memory poisoning is a documented, distinct threat from standard prompt injection: rather than a manipulation that ends when a session closes, an adversary implants false or malicious information into an AI agent's long-term memory storage, and the agent 'learns' the instruction, recalling and acting on it in future sessions — sometimes days or weeks after the original poisoning occurred. This checklist scores your agent setup against the defenses that actually address this persistence: periodic memory review, restricting what untrusted sources can write to memory, memory expiration, and fresh verification for high-impact actions rather than trusting stored memory alone.

• →Addresses memory poisoning specifically, distinct from standard prompt injection which ends with the session
• →Weights restricting untrusted sources from writing to long-term memory as a top defense
• →Covers memory expiration and scoped storage, both of which limit a poisoned entry's blast radius and lifespan
• →Includes fresh-verification requirements for high-impact actions rather than trusting persisted memory
• →Watermarked by Cikal Studio Labs · Works on any device, no install required