AI agents that act on emails or documents can be tricked by hidden instructions. Without memory the trick only works once. With memory a harmful instruction can be saved and followed in future sessions turning it into a persistent rule.
MemoryDojo is a new that tests this risk. It runs agents on long sequences of tasks injects a poisoned instruction at one point and tracks whether memory preserves and acts on it later.
Every memory system tested was vulnerable. An optimized attack increased effectiveness by 2% to 31% depending on architecture. This shows memory design is a critical security surface.