Memory Statue v2: A Structured Memory Architecture for Long-Horizon AI Agents

Key Points

• To develop a structured memory architecture that enhances the performance and reliability of long-horizon AI agents.
• Introduced a structured memory store called Memory Crystal with organized capsules and indices.
• Developed deterministic retrieval pipelines integrated with policy gates for better information retrieval.
• Implemented mechanisms for promoting canonical beliefs and managing memory with append-only audit trails.
• Employed drift modeling techniques, including survival analysis, to monitor memory integrity over time.
• Provided specifications for evaluating memory health metrics, such as contradiction rates and staleness.
• Demonstrated a significant increase in memory reliability compared to existing prompt-based methods.
• Showcased methods for drift reduction through measurable defect metrics and maintenance routines.

Abstract