Summary
AdaMem is an adaptive, user-centric memory framework for long-horizon dialogue agents, developed by Tsinghua University, WeChat, and USTC (arXiv:2603.16496). It addresses three weaknesses of existing systems: over-reliance on semantic vector similarity, fragmented memories caused by chunking, and the granularity dilemma of chunk sizing. AdaMem replaces flat vector stores with a four-layer structure: Working Memory (recent dialogue, FIFO queue), Episodic Memory (structured event diaries and facts), Persona Memory (cross-conversation user profiles), and a lateral Graph Memory that indexes relationships across time and episodes. Retrieval is adaptive and user-centric: it first identifies whether a question targets the user or the assistant, routes queries by rule (when/why to graph expansion, who to persona lookup), then fuses evidence with roughly 70% semantic and 30% graph-based retrieval. A multi-agent pipeline (Memory Agent, Research Agent, Working Agent) maintains memory, retrieves evidence, and generates answers. Experiments show a 23.4% gain in temporal reasoning (LoCoMo benchmark, F1 44.65) and 63.25% user-profile accuracy on PersonaMem with +27.3% generalization; ablations identify Graph Memory as the largest contributor.
AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents
- Authors/Affiliations: Tsinghua University, WeChat, University of Science and Technology of China
- Paper: arxiv.org/pdf/2603.16496
Key points
Motivation: three pain points of existing memory systems
- Over-reliance on semantic similarity: vector matching struggles to surface evidence that is not literally similar to the query.
- Fragmented memory: chunk splitting breaks temporal, causal, and character relationships.
- Granularity dilemma: coarse chunks introduce noise; fine chunks lose the macro context.
Core design: four-layer structured memory
Instead of a flat vector store, AdaMem uses vertical hierarchical consolidation plus a horizontal relational index:
- Working Memory — a "desktop sticky note" of recent verbatim dialogue, managed as a FIFO queue.
- Episodic Memory — event diaries with facts and attributes, extracted in structured form.
- Persona Memory — user profiles aggregated across conversations.
- Graph Memory — a lateral relational board connecting clues across time and episodes.
Adaptive retrieval strategy
1. Target identification: determine whether the question refers to the User or the Assistant, resolving ambiguous references.
2. Route planning (rules first): when/why questions go to graph expansion; who questions go to persona lookup.
3. Evidence fusion: semantic retrieval (~70%) combined with graph-relation completion (~30%).Experimental highlights
- Temporal reasoning: +23.4% improvement on the LoCoMo benchmark (F1 Score 44.65).
- User profiling: 63.25% accuracy on PersonaMem, with +27.3% generalization.
- Key finding: ablation studies show Graph Memory is the most critical source of gains — cross-time/cross-episode relation completion effectively solves the fragmentation problem.
Multi-agent collaboration
The framework operates via three cooperating agents:
1.
Memory Agent — maintains and updates memory.
2.
Research Agent — retrieves evidence.
3.
Working Agent — generates responses.
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177168996