Research map / Architectures and efficiency / Attention and transformers
CoMem: Context Management with A Decoupled Long-Context Model
Yuwei Zhang, Chengyu Dong, Shuowei Jin, Changlong Yu, Hejie Cui, Hongye Jin, Xinyang Zhang, Hamed Bonab, Colin Lockard, Jianshu Chen, Zhenyu Shi, Jingbo Shang and 2 more
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 |
|---|---|---|---|
| add GRPO functional-equivalence training to the memory model (vs base Qwen3-4B) (ablation) | CoMem · SWE-Bench-Verified, GLM-4.7 agent | % resolved: 58.3 -> 62.7 | Improved |
| add GRPO training to the memory model (vs base Qwen3-4B) (ablation) | CoMem · SWE-Bench-Verified, DeepSWE-32B agent | % resolved: 29.9 -> 41.0 | Improved |
| add GRPO training to the memory model (vs base Qwen3-4B) (ablation) | CoMem · SWE-Bench-Verified, Qwen3-Coder-Max agent | % resolved: 42.2 -> 51.0 | Improved |
| add GRPO training on top of SFT memory model (ablation) | CoMem · SWE-Bench-Verified, Qwen3-Coder-Max agent | % resolved: 47.3 -> 51.0 | Improved |
| add a summary from the memory model (vs No Summary) (ablation) | CoMem · SWE-Bench-Verified, GLM-4.7 agent | % resolved: 59.4 -> 62.7 | Improved |
| add a summary from the memory model (vs No Summary) (ablation) | CoMem · SWE-Bench-Verified, Qwen3-Coder-Max agent | % resolved: 46.6 -> 51.0 | Improved |
| replace full-context inference with CoMem | DeepSWE-32B · SWE-Bench-Verified | % resolved: 40.4 -> 41.0 | No clear effect |
| replace full-context inference with CoMem | GLM-4.7 · SWE-Bench-Verified | % resolved: 69.0 -> 62.7 | Hurt |
| replace full-context inference with CoMem | GLM-4.7 · SWE-Bench-Verified | speedup (CPU offload): 1× -> 2.08× | Improved |
| replace MemAgent 7B memory with CoMem GRPO memory | CoMem · SWE-Bench-Verified, GLM-4.7 agent | % resolved: 54.8 -> 62.7 | Improved |
| replace full-context with CoMem on BrowseComp-EN | GLM-4.7 · BrowseComp-EN | accuracy: 28.1 -> 32.0 | Improved |
| set update frequency k=2 instead of k=4 (ablation) | CoMem · SWE-Bench-Verified, GLM-4.7 agent | % resolved: 62.7 -> 64.2 | No clear effect |
Closest papers by meaning
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.90)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.89)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.88)
- MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents (ICLR 2026 · similarity 0.88)
- A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.88)
- Scaling Long-Horizon Agent via Context Folding (ICML 2026 · similarity 0.88)
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.
