Survey Examines Security Risks in LLM Agents' Long-Term Memory Systems

A research survey addresses security vulnerabilities in large language model agents with persistent memory systems. The study, published as arXiv:2604.16548v1, shifts focus from traditional data leakage concerns to more complex threats involving continuous shaping and unauthorized access across sessions. Drawing from cognitive neuroscience and philosophy of memory, the analysis characterizes agent memory as malleable, rewritable, and socially propagating. Researchers developed a memory-lifecycle framework organized around six phases—Write, Store, Retrieve, Execute, Share, and Forget/Rollback—cross-tabulated against four security objectives: integrity, confidentiality, availability, and governance. The survey identifies persistent, writable memory as an independent security problem distinct from previous architectural concerns. Recent research has primarily examined memory architectures and agent mechanisms, but this work centers on the epistemic and governance properties that create unique vulnerabilities. The concept of mnemonic sovereignty emerges as a central concern in protecting agent memory systems from cross-session poisoning and propagation across shared organizational states.