AdaMem: Adaptive User-Centric Memory Organizes Long-Horizon Dialogue Agents into Four Tiers

arXiv 2603.16496·medium signal

AdaMem introduces a four-tier memory architecture for long-horizon dialogue agents—working, episodic, persona, and graph memory—enabling agents to maintain recent context, structured long-term experience, stable user traits, and relation-aware connections in a unified framework. The adaptive routing decides which tier to write/read based on the nature of each interaction. This is a concrete, implementable memory architecture for agent builders dealing with context window limits in persistent conversational deployments.