Three-Layer Agent Memory

Problem Solved

Agents often keep important conversations, user preferences, and events scattered across context, making later retrieval, deduplication, and retention decisions difficult. Three-Layer Agent Memory is designed for AI agent memory management. It classifies dialogue into waste, profile, and event, then organizes event memory into L1, L2, and L3 pages: L1 for fast lookup, L2 for summaries, and L3 for the full original content.

How It Works

• Write and retrieve: memory_process processes dialogue and stores it with classification; memory_search and memory_read locate historical events and layered content.
• User profile: memory_profile_get and memory_profile_update maintain structured preferences, habits, promises, and commitments; updates should generally add rather than delete history.
• Merge and archive: memory_detect_similar finds similar pages, suggesting merges at higher confidence thresholds; memory_merge performs the merge, while memory_auto_archive archives old pages by age.
• Daily maintenance: A scheduled task can summarize the day's memory, detect similar entries, archive pages beyond a configured age, and write important events back into the memory store.

Boundaries

It suits long-term assistant, customer support, and profile-driven agent systems, but it relies on scheduled jobs or explicit calls to trigger maintenance. Content marked waste is not stored, and merging or archiving changes what is actively available. Production deployments should keep logs and a recovery strategy.