The paper formalizes long-term agent memory as a governed evolving state with correctness defined over trajectories rather than records (arXiv:2605.26252). Four failure modes recur across storage models: unregulated growth, missing semantic revision, capacity-driven forgetting, and read-only retrieval. No record-level system satisfies the six state-evolution conditions regardless of backend. MemState implements the four operators on a property-graph store and exposes gaps to a native engine. Related systems such as MemGPT and vector-augmented agent frameworks exhibit the same record-centric limits on revision and forgetting. Three structural observations establish the separation between current database workloads and the required memory-centric workload.