MemoBrain: Executive Memory as an Agentic Brain for Reasoning
Hongjin Qian, Zhao Cao, Zheng Liu
2026-01-14
Summary
This paper introduces MemoBrain, a new system designed to help AI agents think through complex tasks that require many steps, especially when those agents use tools like web browsers or search engines.
What's the problem?
When AI agents try to solve complicated problems, they need to remember what they've already done and thought. However, large language models, which power these agents, have limited 'short-term memory'. As the agent works, its reasoning process and the results from using tools fill up this memory, making it hard to keep track of the overall goal and leading to mistakes or getting lost in the process. It's not just about making things faster, but about maintaining a clear line of thought.
What's the solution?
MemoBrain acts like a 'co-pilot' for the AI agent. It doesn't directly control the agent, but it organizes the agent's thinking process. It does this by identifying the most important steps, getting rid of irrelevant information, and summarizing completed parts of the task. Essentially, it creates a condensed, logical 'backbone' of the reasoning process, ensuring the agent stays focused and doesn't exceed its memory limits. It actively manages what the agent remembers, rather than just letting information pile up.
Why it matters?
This work is important because it shows that memory isn't just a helpful addition to AI agents, it's crucial for them to be able to handle complex, multi-step tasks effectively. By giving agents a way to manage their reasoning process and maintain a coherent train of thought, MemoBrain allows them to tackle more challenging problems and achieve better results, as demonstrated on several difficult benchmarks.
Abstract
Complex reasoning in tool-augmented agent frameworks is inherently long-horizon, causing reasoning traces and transient tool artifacts to accumulate and strain the bounded working context of large language models. Without explicit memory mechanisms, such accumulation disrupts logical continuity and undermines task alignment. This positions memory not as an auxiliary efficiency concern, but as a core component for sustaining coherent, goal-directed reasoning over long horizons. We propose MemoBrain, an executive memory model for tool-augmented agents that constructs a dependency-aware memory over reasoning steps, capturing salient intermediate states and their logical relations. Operating as a co-pilot alongside the reasoning agent, MemoBrain organizes reasoning progress without blocking execution and actively manages the working context. Specifically, it prunes invalid steps, folds completed sub-trajectories, and preserves a compact, high-salience reasoning backbone under a fixed context budget. Together, these mechanisms enable explicit cognitive control over reasoning trajectories rather than passive context accumulation. We evaluate MemoBrain on challenging long-horizon benchmarks, including GAIA, WebWalker, and BrowseComp-Plus, demonstrating consistent improvements over strong baselines.