Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MemoBrain: Executive Memory as an Agentic Brain for Reasoning

Hongjin Qian, Zhao Cao, Zheng Liu

TL;DR

The paper addresses the challenge of sustaining coherent, long-horizon reasoning in tool-augmented LLM agents within bounded contexts. It introduces MemoBrain, an executive memory system that constructs a dependency-aware memory graph from reasoning episodes and actively manages the working context through folding and flushing under a fixed budget, operating asynchronously as a copilot to the reasoning agent. The approach is trained in two stages—memory construction via supervised fine-tuning and memory management via direct preference optimization—and shows consistent improvements across GAIA, WebWalker, and BrowseComp-Plus benchmarks, with the best results when paired with larger tool-augmented bases like DeepResearch-30B-A3B. This work demonstrates that explicit cognitive control over memory and information flow can substantially improve task performance and scalability of long-horizon reasoning systems in real-world information-seeking tasks, enabling more robust and scalable agent behavior under tight context budgets. $32K$ and $64K$ token budgets are central design choices, illustrating the practical bounds within which MemoBrain operates effectively.

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.

MemoBrain: Executive Memory as an Agentic Brain for Reasoning

TL;DR

The paper addresses the challenge of sustaining coherent, long-horizon reasoning in tool-augmented LLM agents within bounded contexts. It introduces MemoBrain, an executive memory system that constructs a dependency-aware memory graph from reasoning episodes and actively manages the working context through folding and flushing under a fixed budget, operating asynchronously as a copilot to the reasoning agent. The approach is trained in two stages—memory construction via supervised fine-tuning and memory management via direct preference optimization—and shows consistent improvements across GAIA, WebWalker, and BrowseComp-Plus benchmarks, with the best results when paired with larger tool-augmented bases like DeepResearch-30B-A3B. This work demonstrates that explicit cognitive control over memory and information flow can substantially improve task performance and scalability of long-horizon reasoning systems in real-world information-seeking tasks, enabling more robust and scalable agent behavior under tight context budgets. and token budgets are central design choices, illustrating the practical bounds within which MemoBrain operates effectively.

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.
Paper Structure (31 sections, 12 equations, 3 figures, 7 tables)

This paper contains 31 sections, 12 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Overview of MemoBrain
  5. Memory Construction.
  6. Memory Management.
  7. Memory Construction
  8. Thought Formation.
  9. Dependency Modeling.
  10. Memory Management
  11. Sequential Trajectory Folding.
  12. Selective Memory Flush.
  13. Context Reorganization.
  14. Optimization
  15. Stage I: Memory Construction.
  16. ...and 16 more sections

Figures (3)

  • Figure 1: MemoBrain asynchronously abstracts transient reasoning sessions into compact, dependency-aware memory units and maintains a global reasoning memory. Upon reaching the memory budget, it constructs a memory-managed context by folding or flushing past information, enabling coherent long-horizon reasoning.
  • Figure 2: Ablation studies of MemoBrain. All experiments are conducted on 100 randomly sampled instances from BrowseComp-Plus, using DeepResearch-30B-A3B as the base reasoning agent.
  • Figure 3: Efficiency Analysis of MemoBrain. End-to-end runtime statistics on the full BrowseComp-Plus benchmark under a 32K memory budget, using 4×H100 GPUs for the reasoning model and 2×H100 GPUs for the memory model with a batch size of 8.