Researchers Show a Single Email Can Poison an AI Assistant's Long-Term Memory

A newly published study demonstrates that a single crafted email can silently plant false information into the persistent memory of a personal AI agent, altering how it responds and acts in later sessions without the user ever noticing. The technique, dubbed MemGhost, targets a rapidly growing class of always-on assistants that combine long-term memory with access to inboxes, calendars and file systems.