Long-Term Memory for a Conversational Assistant

Abstract

Most AI assistants either remember nothing or remember far too much. This paper introduces ATAPIC's long-term memory architecture, a system that distills conversations into durable facts instead of storing raw transcripts. By combining intelligent memory extraction, semantic retrieval, recency-based recall, and complete user ownership, the memory layer enables assistants that become more useful over time while remaining private, explainable, and fully controllable. The paper explores the complete architecture, storage model, retrieval pipeline, governance model, and how long-term memory powers personalized conversations, Plays, and autonomous AI agents.