Research map / Large language models / Retrieval and knowledge
TransMem: Transition-Aware Retrieval for Evolving Personal Memory
Shigeng Chen, Linhao Luo, Changlong Shi, Zhangchi Qiu, Shirui Pan
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TransMem, a transition-aware retrieval approach for evolving personal memory that links how earlier preferences are refined, contradicted, or superseded by later evidence.
Problem
Long-term personalized agents must answer queries over user histories where preferences, goals, and constraints evolve across interleaved conversations; existing retrieval and structured-memory systems return isolated memory units without explaining how earlier preferences changed.
Approach
Retrieval that is aware of transitions between memory states (refinement, contradiction, supersession) so that returned memory explains which earlier preference was updated by later evidence, rather than isolated memory units.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add transition records and adaptive expansion over factual anchors only (ablation) | TransMem · PersonaMem (4 settings) | overall accuracy | Improved |
| add intra-topic transition edges to factual anchors (ablation) | TransMem · PersonaMem-32k Qwen3-8B | overall accuracy (pp): +0.42 pp | Improved |
| add inter-topic transition edges to factual anchors (ablation) | TransMem · PersonaMem-32k Qwen3-8B | overall accuracy (pp): +1.05 pp | Improved |
| use both intra and inter-topic edges instead of one type (ablation) | TransMem · PersonaMem (4 settings) | overall accuracy | Improved |
| use graph-grounded transition expansion instead of chronological neighbors (ablation) | TransMem · PersonaMem-32k / 128k Qwen3-8B | overall accuracy: 53.47 -> 56.84 / 46.51 -> 48.03 | Improved |
| use graph expansion instead of flat transition retrieval (ablation) | TransMem · PersonaMem-32k / 128k Qwen3-8B | overall accuracy: 54.32 -> 56.84 / 46.92 -> 48.03 | Improved |
| use adaptive stopping instead of best fixed hop depth (ablation) | TransMem · PersonaMem-32k / 128k Qwen3-8B | overall accuracy: +1.68 pp / +0.83 pp | Improved |
| add transition context for New Ideas questions (ablation) | TransMem · PersonaMem-128k Qwen3-8B | New Ideas accuracy: -3.27 pp | Hurt |
| add transition records for Aligned Recommendations (ablation) | TransMem · PersonaMem-128k Qwen3-8B | Aligned Recs accuracy: unchanged | No clear effect |
| compare retrieval token cost with A-Mem | TransMem · PersonaMem-32k Qwen3-8B | inference tokens / latency: 3.4k -> 1.82k tokens; 2.35 s -> 0.98 s | Improved |
Experimental setup
Baselines: retrieval-based memory systems structured-memory systems
Closest papers by meaning
- Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.90)
- AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations (ICLR 2026 · similarity 0.88)
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (ICLR 2025 · similarity 0.87)
- SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents (ICLR 2025 · similarity 0.87)
- Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation (NeurIPS 2026 · similarity 0.87)
- Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
