Memory Statue v2 proposes a structured memory architecture for long-horizon AI agents that addresses limitations of LLM prompt-based memory approaches. The architecture features a Memory Crystal (structured memory store with capsules and indices), deterministic retrieval pipelines with policy gates, governed promotion of canonical beliefs, append-only audit trails, and drift reduction via measurable defect metrics and maintenance routines. The paper presents system specifications including data schemas, retrieval logic, drift modeling with survival analysis, and a validation harness for measuring memory health metrics such as contradiction rates, unfaithfulness, and staleness. This is a system specification and research blueprint, emphasizing testable, governable memory that is resistant to drift over extended operational periods.
Lloyd D. Frazier (Thu,) studied this question.