CoMem: Efficient Agent Memory Management via Decoupling Long-Context Models

CoMem is a new context management framework whose core lies in decoupling memory management from the main agent workflow and executing it asynchronously, significantly reducing response latency for long-context tasks while maintaining performance. Its key designs include the k-step offset asynchronous pipeline strategy and reward-driven memory alignment training, achieving a 1.4x latency improvement on the SWE-Bench-Verified benchmark and providing a new path for modular optimization of agent systems.

Modern agents handle complex tasks by iteratively summarizing historical interactions, but each summary token generation introduces additional decoding overhead, which translates to end-to-end response latency and severely impacts user experience (e.g., the waiting issue when a programming assistant reviews conversation history). This is the core dilemma of current context management methods.

CoMem fully decouples memory management from the main agent workflow and adopts the "k-step offset asynchronous pipeline" strategy: the memory model continuously summarizes historical interactions in the background, while the main agent focuses on current reasoning and retrieves the latest completed summary (which may be slightly outdated) when accessing memory. The k value needs to balance update timeliness and system overhead, and the optimal solution is found through theoretical analysis and experiments.

To ensure that memory summaries are useful for decision-making in asynchronous scenarios, CoMem uses reward-driven training: it evaluates the contribution of memory summaries to the quality of agent decisions, converts this into reward signals to guide the memory model's learning, enabling it to not only compress information but also retain key statistical information for decision-making, thus ensuring the effectiveness of reasoning in asynchronous scenarios.

In the SWE-Bench-Verified benchmark test, CoMem achieves a 1.4x latency improvement compared to traditional long-context solutions, while the performance degradation is mild. The information lag introduced by asynchrony is effectively mitigated through reward training, and in most cases, the agent can still make correct decisions based on slightly outdated memory.

CoMem's decoupled architecture provides a new idea for modular optimization of agent systems: it allows independent improvement of memory compression and reasoning strategies without worrying about mutual interference. This framework can naturally be extended to support diverse memory types (such as episodic, semantic, and procedural memory), helping to expand agent application scenarios.

CoMem currently has limitations: fixed k-step offset (can be dynamically adjusted in the future), only supports text interaction (needs to expand to multimodality), and task-agnostic memory model (can customize task-specific models). These directions are the focus of future optimization.