The more AI remembers about you, the longer one mistake can follow you. It turns out an assistant that records a deadline, preference, financial detail, or project decision incorrectly may keep using that false information across future conversations. Unlike a one-time hallucination, the mistake becomes part of what the system thinks it knows about you. A new preprint tested this problem by poisoning just 1.2% of an agent’s memory corpus. Answer accuracy fell from 85% to 30%. The false memories looked like ordinary facts. One simply changed the number of Spotify playlists a user had from 20 to 15. A prompt-injection screener caught 83.2% of indirect injections but none of the 360 poisoned memories. That's the problem with defending agent memory. Malicious instructions often look suspicious but a false statement may look completely normal. Provenance can't prove that a memory is true but it can preserve where the information came from, who introduced it, and why the agent trusted it. As AI assistants remember more of our lives, correcting what they know will matter just as much as expanding what they can remember.(Sahara AI 🔆)