Research map / Large language models / Retrieval and knowledge
The Pensieve Paradigm: Stateful Language Models with Learned Memory Management
Xiaoyuan Liu, Tian Liang, Dongyang Ma, Deyu Zhou, Haitao Mi, Pinjia He, Yan Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train self-context management (StateLM) instead of plain instruct model | Qwen3-8B · NovelQA | accuracy: 65.87 -> 83.84 | Improved |
| train self-context management (StateLM) | Qwen3-8B · LongMemEval-S | accuracy: 45.40 -> 58.93 | Improved |
| train self-context management (StateLM) | Qwen3-14B · BrowseComp-Plus | accuracy: 5.46 -> 51.33 | Improved |
| train self-context management (StateLM) | Qwen3-4B · BrowseComp-Plus | accuracy: 2.89 -> 35.33 | Improved |
| add reinforcement learning on top of SFT StateLM (ablation) | StateLM-8B · ∞Bench | accuracy: 70.16 -> 73.07 | Improved |
| add reinforcement learning on top of SFT StateLM (ablation) | StateLM-14B · BrowseComp-Plus | accuracy: 51.33 -> 52.67 | Improved |
| add reinforcement learning on top of SFT StateLM (ablation) | StateLM-8B · BrowseComp-Plus | accuracy: 46.22 -> 46.44 | No clear effect |
| continue SFT with additional epochs instead of RL (ablation) | StateLM · long-context benchmarks | accuracy | Hurt |
| use Qwen3 instruct with agentic prompt and StateLM toolkit but no training (ablation) | Qwen3-14B-Agentic · NovelQA and ∞Bench | completion rate: up to 30% of questions completed | Hurt |
| prune only the assistant tool-call segment in deletion for NIAH (ablation) | StateLM · NIAH | accuracy | Improved |
| remove search tool for NIAH (ablation) | StateLM · NIAH | accuracy | Hurt |
| use StateLM vs Qwen3 instruct at 2M NIAH | StateLM-14B · NIAH 2M | accuracy: 1.7 -> 83.89 | Improved |
Closest papers by meaning
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.88)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.87)
- Minerva: A Programmable Memory Test Benchmark for Language Models (ICML 2025 · similarity 0.87)
- A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.87)
- AUTOMEM: Automated Learning of Memory as a Cognitive Skill (NeurIPS 2026 · similarity 0.87)
- REMem: Reasoning with Episodic Memory in Language Agent (ICLR 2026 · similarity 0.87)
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.
