Saved in:
Bibliographic Details
Main Authors: Zhang, Yuwei, Dong, Chengyu, Jin, Shuowei, Yu, Changlong, Cui, Hejie, Jin, Hongye, Zhang, Xinyang, Bonab, Hamed, Lockard, Colin, Chen, Jianshu, Shi, Zhenyu, Shang, Jingbo, Li, Xian, Yin, Bing
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.30842
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910271654592512
author Zhang, Yuwei
Dong, Chengyu
Jin, Shuowei
Yu, Changlong
Cui, Hejie
Jin, Hongye
Zhang, Xinyang
Bonab, Hamed
Lockard, Colin
Chen, Jianshu
Shi, Zhenyu
Shang, Jingbo
Li, Xian
Yin, Bing
author_facet Zhang, Yuwei
Dong, Chengyu
Jin, Shuowei
Yu, Changlong
Cui, Hejie
Jin, Hongye
Zhang, Xinyang
Bonab, Hamed
Lockard, Colin
Chen, Jianshu
Shi, Zhenyu
Shang, Jingbo
Li, Xian
Yin, Bing
contents Context management enables agentic models to solve long-horizon tasks through iterative summarization of previous interaction histories. However, this process typically incurs substantial decoding overhead for the extra summarization tokens, which significantly affect the end-to-end response latency at deployment. In this paper, we introduce CoMem, a novel framework that decouples memory management from the primary agent workflow, enabling these processes to execute in parallel. We propose a $k$-step-off asynchronous pipeline that overlaps the memory model's summarization with the agent's inference, effectively masking the latency of context processing. To ensure robustness under this asynchronous setting, we introduce a reward-driven training strategy that aligns the memory model to capture sufficient statistics for the agent's decision-making. Theoretical analysis confirms that CoMem offers a superior efficiency-effectiveness trade-off compared to coupled architectures. Our extensive experimental results on SWE-Bench-Verified show that CoMem provides 1.4x latency improvements upon vanilla long-context solutions while preserving most of the performance. Furthermore, we demonstrate that these latency gains scale favorably with increased system throughput, offering a modular path forward for the independent optimization of agent reasoning and memory compression.
format Preprint
id arxiv_https___arxiv_org_abs_2605_30842
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle CoMem: Context Management with A Decoupled Long-Context Model
Zhang, Yuwei
Dong, Chengyu
Jin, Shuowei
Yu, Changlong
Cui, Hejie
Jin, Hongye
Zhang, Xinyang
Bonab, Hamed
Lockard, Colin
Chen, Jianshu
Shi, Zhenyu
Shang, Jingbo
Li, Xian
Yin, Bing
Machine Learning
Context management enables agentic models to solve long-horizon tasks through iterative summarization of previous interaction histories. However, this process typically incurs substantial decoding overhead for the extra summarization tokens, which significantly affect the end-to-end response latency at deployment. In this paper, we introduce CoMem, a novel framework that decouples memory management from the primary agent workflow, enabling these processes to execute in parallel. We propose a $k$-step-off asynchronous pipeline that overlaps the memory model's summarization with the agent's inference, effectively masking the latency of context processing. To ensure robustness under this asynchronous setting, we introduce a reward-driven training strategy that aligns the memory model to capture sufficient statistics for the agent's decision-making. Theoretical analysis confirms that CoMem offers a superior efficiency-effectiveness trade-off compared to coupled architectures. Our extensive experimental results on SWE-Bench-Verified show that CoMem provides 1.4x latency improvements upon vanilla long-context solutions while preserving most of the performance. Furthermore, we demonstrate that these latency gains scale favorably with increased system throughput, offering a modular path forward for the independent optimization of agent reasoning and memory compression.
title CoMem: Context Management with A Decoupled Long-Context Model
topic Machine Learning
url https://arxiv.org/abs/2605.30842