AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents

Forum topic · ✨步子哥 · 2026-03-22

Summary

AdaMem is an adaptive, user-centric memory architecture for long-horizon dialogue agents, developed by researchers from Tsinghua University, WeChat, and USTC. It addresses three weaknesses of flat vector-store memory: over-reliance on semantic similarity, fragmentation caused by chunking (breaking temporal, causal, and interpersonal links), and difficulty choosing the right granularity. AdaMem organizes memory in four stores: a working memory (recent raw dialogue, FIFO), an episodic memory (structured event diaries and facts), a persona memory (aggregated user profile across conversations), and a graph memory that indexes relations across layers. Retrieval is adaptive: it first identifies whether a question targets the user or the assistant, routes queries by rule (when/why via graph expansion, who via persona lookup), then fuses evidence weighted 70% semantic and 30% graph relations. On the LoCoMo benchmark it reports an F1 of 44.65 and a 23.4% improvement in temporal reasoning; on PersonaMem it achieves 63.25% persona accuracy with 27.3% better generalization. Ablation studies identify graph memory as the largest contributor. The system uses three cooperating agents: Memory Agent (maintenance), Research Agent (evidence retrieval), and Working Agent (answer generation). Paper: arxiv.org/pdf/2603.16496

AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents — from Tsinghua University, WeChat, and USTC (arxiv.org/pdf/2603.16496)

Motivation: Three Pain Points in Existing Memory Systems

Semantic dependency: over-reliance on vector similarity misses deep evidence that is not a literal match.
Memory fragmentation: chunking breaks temporal, causal, and interpersonal relationship chains.
Granularity dilemma: coarse splits introduce noise; fine splits lose the macro narrative.

Core Innovation: Four-Store Memory Structure

Vertical layered consolidation plus a horizontal relational index replaces a flat vector database:

Working Memory — sticky-note style recent dialogue in raw form, FIFO queue.
Episodic Memory — event diaries; facts and attributes extracted in structured form.
Persona Memory — character profile / user portrait aggregated across conversations.
Graph Memory — a relational clue board and association index network spanning the layers.

Adaptive Retrieval Strategy

1. Target identification — determine whether the question targets the User or the Assistant; resolve ambiguous references. 2. Route planning — rules first: *when/why* → graph expansion; *who* → persona lookup. 3. Evidence fusion — semantic-dominant (70%) + graph-relation completion (30%).

Experimental Highlights

23.4% improvement in temporal reasoning — LoCoMo benchmark, F1 Score 44.65.
63.25% user-persona accuracy — PersonaMem benchmark, with +27.3% generalization.
Key finding: ablation studies show Graph Memory is the single most important source of gains; cross-time/cross-chunk relational completion effectively solves the fragmentation problem.

Multi-Agent Collaboration

Memory Agent — maintains memory.
Research Agent — retrieves evidence.
Working Agent — generates answers.