AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent

Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, Wei-Ying Ma, Jingjing Liu, Mingxuan Wang, Hao Zhou

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs

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 changedWhereMeasuredEffect
train memory agent with multi-conv DAPO RL instead of using MemAgent workflow untrained (ablation)MemAgent · RULER-HQA and LongBench-QAaccuracyImproved
apply MemAgent workflow without RL (ablation)MemAgent · LongBench-QAaccuracyNo clear effect
MemAgent without RL versus backbone (ablation)MemAgent · RULER-HQAaccuracyMixed
process text in chunks with fixed-length overwrite memory (RL-MemAgent-14B)Qwen2.5-14B-Instruct · RULER-HQA 3.5Maccuracy: 80.47 (7K) -> 71.09 (3.5M)Improved
process text in chunks with memory vs Qwen2.5-Instruct-1MQwen2.5-Instruct-14B-1M · RULER-HQA 896Kaccuracy: 0.00 -> 75.78Improved
RL-MemAgent on LongBench-QA vs larger long-context modelsQwen2.5-14B-Instruct · LongBench-QA averagescore: 50.7 (QwenLong-L1-32B) -> 51.0Improved
RL-MemAgent for summarizationQwen2.5-14B-Instruct · LongBench-SUM GovReportROUGE average: 19.34 -> 21.80Improved
larger memory (4096) vs default memory 1024 (ablation)MemAgent-14B · LongBench-QAaverage score: 51.0 (m1024) vs 48.5 (m4096)No clear effect
memory length 256 to 4096 on NIAH (ablation)MemAgent · NIAHaccuracy: 97.92 to 98.18 at 512K (14B)No clear effect
shift key information position in context (probe) (ablation)MemAgent-14B · RULER-HQA probeaccuracy: 0.00 to 2.79 avg diff vs randomNo clear effect
use RL-trained memory instead of RAG retrieval (top-K)Qwen2.5-14B · RULER-HQA 3.5Maccuracy: 66.41 -> 71.09Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy