Paper rethinks AI agent memory as state trajectory, not storage

A new arXiv paper challenges the fundamental assumption that long-running AI agent memory should be built on traditional database paradigms, such as records, vector embeddings, or knowledge graphs. Authors Orogat and Mansour identify four recurring failure modes in current systems: unregulated memory growth (no built-in forgetting), missing semantic revision (cannot update past beliefs), capacity-driven forgetting (drops old data without semantic judgment), and read-only retrieval (no ability to mutate stored state). They argue these failures are inherent because today's systems treat memory as static storage where correctness is localized to individual records, embeddings, or edges.

“Long-term agent memory is a new data-management workload,” the authors propose, introducing “Governed Evolving Memory” (GEM). GEM redefines memory correctness as a property of the entire state trajectory rather than individual records. It replaces record-level operations with four state-level operators: ingestion (add new info), revision (modify existing beliefs), forgetting (prune semantically), and retrieval (query state). Six formal correctness conditions (e.g., monotonicity, revision consistency) govern how the state evolves. The paper proves three structural observations—no record-level system can satisfy these conditions, regardless of storage model. The authors implement MemState, a prototype on a property-graph backend, to validate feasibility and identify the gap to a native engine. They outline three research directions for memory-centric data management as a distinct workload, aiming to enable agents that genuinely learn and adapt across long-running sessions.