Long-conversation memory stops searching everything: an agent drills down on demand and reasons over just 8% of dialogues
Synopsis
Personalized assistants must recover sparse evidence from long conversation histories, while queries vary widely in complexity. APDMem organizes history into four layers, from thematic summaries and personalized key facts down to turn-level evidence notes and raw messages, and a controller reads high-level summaries first, drilling into finer evidence only when needed. On LongMemEval it achieves strong long-context memory reasoning while accessing only 8% of the total conversations.
Interpretation
Long-conversation memory retrieval moves from a fixed granularity to controller-driven drilling down through layers, so simple queries can terminate early and only complex ones inspect finer evidence. Earlier approaches relied on a flat memory store or a fixed retrieval granularity, applying one retrieval pattern to every query. On LongMemEval, this layered retrieval achieves strong long-context memory reasoning performance while accessing only 8% of the total conversations.
Conversation history is organized into four progressively detailed layers, from thematic summaries down to raw messages. The hierarchy creates an adjustable trade-off between retrieval cost and answer fidelity instead of pulling back all relevant content at once. Complex temporal, multi-hop, or exact-evidence queries trigger deeper inspection, while simple queries stop at high-level summaries.
Retrieved evidence is first consolidated into a query-focused structure before final answer generation. A note synthesizer merges facts, orders events, and flags contradictions in the evidence. This consolidation step occurs before answer generation and acts on the retrieved evidence.
Perspective
The result targets personalized assistants retrieving evidence from long conversation histories, evaluated on LongMemEval; both the reported performance and the 8% access ratio are confined to that setting. It suits researchers and engineers interested in long-context memory and retrieval cost.
The summary states strong performance on LongMemEval and an 8% access ratio, without listing specific metric values or describing behavior on other datasets or conversation lengths. Layered drill-down triggers deeper inspection for complex queries, and its cost-benefit balance remains tied to this evaluation setting.