CoMem: Context Management with A Decoupled Long-Context Model

2026-05-29

AI Summary

This work addresses the significant decoding latency introduced by iterative summarization in long-horizon tasks, where agents rely on such summaries to manage context. The authors propose CoMem, a novel framework that decouples memory management from the main reasoning pipeline for the first time. CoMem employs a k-step asynchronous pipelined execution to parallelize context summarization and agent decision-making, complemented by a reward-driven memory training mechanism that preserves critical information under asynchrony. Evaluated on SWE-Bench-Verified, CoMem achieves 1.4× lower end-to-end latency compared to conventional long-context approaches while maintaining comparable performance, with its latency advantage further amplifying as system throughput increases.

Abstract

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.

Problem

• context management
• long-horizon tasks
• decoding overhead
• response latency
• agent models

Innovation

• context management
• decoupled architecture
• asynchronous pipeline
• reward-driven training
• long-context modeling