AI Agent Memory Can Be Silently Rewritten to Control Behavior, Research Shows

A researcher demonstrated that an AI agent's associative memory graph can be fabricated or overwritten without any training, reward signals, or lived experience. In experiments, a memory graph authored with zero trials produced the same perfect accuracy (1.00) as one built over 600 rewarded trials, meaning the agent could not distinguish real from fabricated history. The study identified three attack vectors — memory transplant, targeted poisoning, and full history fabrication — all enabled by write-access to the memory store. The key finding reframes agent memory not as data the system reads, but as a bias current that shapes behavior before any decision is made. The researcher argues this demands security standards beyond encryption and signing, treating memory integrity as equivalent to behavioral control.