Source-linked AI summary
MemoBrain: Executive Memory as an Agentic Brain for Reasoning
Hongjin Qian, Zhao Cao, Zheng Liu
TL;DR
Long-horizon tool interactions accumulate transient traces that strain bounded contexts and disrupt task-aligned reasoning. MemoBrain introduces executive memory that builds dependency-aware trajectory memories and actively manages context alongside the agent. It consistently improves performance across challenging benchmarks and model scales, while its benefits depend on agents sustaining sufficiently long reasoning trajectories.
Problem
Long-horizon tool interactions accumulate transient traces and artifacts that strain bounded LLM contexts, impairing logical continuity and task alignment.
Method
MemoBrain is an executive memory copilot that constructs dependency-aware reasoning memories and applies active context operations under a bounded budget.
Results
MemoBrain consistently improves performance across BrowseComp-Plus, GAIA, and WebWalker and across model scales from 4B to 14B.
Takeaways & Limitations
Executive memory treats memory as explicit control over reasoning structure and context rather than passive context storage.
Takeaways & Limitations
MemoBrain has limited opportunity to intervene when reasoning agents terminate prematurely before accumulated context triggers memory management.
Abstract
from arXiv · showhide
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.
1 Introduction
Long-horizon tool use overloads bounded contexts with transient traces, motivating executive memory as active control over reasoning information flow. MemoBrain builds dependency-aware memories and manages context under a fixed budget, improving performance across challenging benchmarks.
- Long-horizon tool interactions accumulate reasoning traces and tool artifacts that strain bounded contexts and impair task alignment.
- Executive memory is task-specific, initialized per task, and evolves online with the reasoning trajectory rather than persisting across tasks.
- MemoBrain operates alongside the reasoning agent as a copilot, delivering task-relevant information just in time while suppressing unnecessary cognitive load.
- MemoBrain abstracts completed reasoning episodes into compact thoughts linked by explicit dependencies, preserving the trajectory’s logical backbone.
- Under a fixed context budget, MemoBrain folds completed sub-trajectories, flushes low-utility elements, and prioritizes high-salience reasoning backbones.
- MemoBrain consistently improves complex information-seeking and reasoning performance across BrowseComp-Plus, GAIA, and WebWalker at model scales from 4B to 14B.
2 Related Work
Related work largely treats memory as persistent knowledge storage or as engineering-oriented context reduction. MemoBrain differs by framing memory as cognitive control over global task awareness and reasoning guidance.
- Long-term memory systems organize documents, retrieved passages, or dialogue histories to support persistent knowledge accumulation and reuse.
- Cross-task and long-term memory commonly support personalization, adaptation, long-context question answering, and interactive assistants.
- AgentFold and ContextFolding reduce context overhead by folding or abstracting redundant reasoning steps.
- DeepAgent delegates subtasks to auxiliary agents and reintegrates summarized outcomes.
- Existing tool-augmented reasoning methods focus on engineering memory rather than cognitive control of global task awareness or reasoning guidance.
3 Method
MemoBrain builds a dependency-aware executive memory alongside the reasoning agent, abstracting episodes into structured thoughts and actively managing context under a bounded budget. It folds resolved sub-trajectories, flushes low-utility steps, and reorganizes the working context while preserving task-relevant reasoning structure.
- Memory Management: When context approaches its budget, MemoBrain predicts executive operations over the trajectory memory and applies them to update the memory state.The updated state is projected into a reorganized working context containing a compact, semantically sufficient representation for subsequent reasoning.
- Memory Construction: MemoBrain abstracts completed reasoning episodes into thoughts and integrates them into a global dependency-aware memory graph.Each thought captures an episode’s semantic contribution, while dependency relations preserve how conclusions build on earlier reasoning steps.
- Memory Construction: MemoBrain operates asynchronously alongside the reasoning agent, continuously structuring progress without interrupting execution.The memory model functions as a copilot rather than blocking the reasoning trajectory.
- Memory Management: FOLD collapses conclusively resolved sub-trajectories into summary thoughts and removes their intermediate steps from the active reasoning context.This operation targets connected sequences of thoughts addressing the same resolved subproblem.
- Memory Management: FLUSH replaces invalid, superseded, or no-longer-relevant thoughts with compact thoughts that preserve structural information while discarding execution-level detail.Unlike folding, flushing does not assume that the associated subtask has been resolved.
- Training: MemoBrain uses one prompt-conditioned memory model in a two-stage strategy that separates semantic abstraction from sparse, decision-driven memory management.Stage I uses supervised fine-tuning with teacher-generated thought annotations, while Stage II predicts operations under a bounded context budget.
4 Experiment
MemoBrain is evaluated on GAIA, WebWalker, and BrowseComp-Plus against diverse reasoning and context-management baselines. Results show consistent gains, with structured memory operations, targeted optimization, and practical efficiency supporting long-horizon reasoning.
- Main Results: MemoBrain improves both GLM-4.6 and DeepResearch-30B-A3B across all evaluated benchmarks, with DeepResearch-30B-A3B achieving the best overall performance.MemoBrain functions as a plug-in memory module without architecture-specific customization.
- Main Results: MemoBrain delivers larger gains on harder regimes such as BrowseComp-Plus and GAIA Level 3, where long-horizon tool interaction strains bounded context.Augmented agents also make more valid tool calls under bounded-context constraints.
- Main Results: Compared with DeepAgent and AgentFold, MemoBrain externally decouples memory control from reasoning execution through a reusable copilot-style design.The paper reports stronger empirical performance and greater flexibility across tool-augmented frameworks.
- Ablation Study: Removing major components degrades performance, while folding or selective flushing outperform unstructured deletion and random step removal.The ablation indicates that gains depend on structured, semantics-aware operations preserving the logical backbone, not merely shorter contexts.
- Ablation Study: Executive memory improves performance across Qwen3 4B, 8B, and 14B backbones, while targeted optimization matters more than model scale alone.MemoBrain-8B is adopted in most experiments as an efficiency-effectiveness trade-off.
- Ablation Study: Below 16K, the memory budget can underperform the baseline, whereas gains begin saturating beyond 64K.Across settings, larger budgets generally improve performance, and the system remains robust within a reasonable range.
- Efficiency Analysis: For trajectories up to 128K tokens, asynchronous memorization and management overheads are negligible relative to reasoning time.MemoBrain also supports reasoning over hundreds of thousand tokens and more tool calls under a smaller fixed context budget.
5 Conclusion
The paper frames executive memory as explicit control over reasoning structure for long-horizon tool-augmented agents. MemoBrain combines a global reasoning graph with budget-aware folding and flushing, and experiments report consistent performance gains alongside validated efficiency and scalability.
- Conclusion: MemoBrain addresses accumulated reasoning traces and tool artifacts that strain bounded contexts and impair task alignment and coherence.The paper treats memory as a control mechanism rather than passive context storage.
- Conclusion: MemoBrain maintains a global reasoning graph and applies budget-aware folding and flushing to provide compact, just-in-time context.The design is presented as a copilot-style memory system for tool-augmented reasoning.
- Conclusion: Experiments on multiple challenging benchmarks demonstrate consistent and substantial performance gains, while ablation and efficiency studies validate the design’s effectiveness and scalability.Additional case studies are provided in Appendix A.3.
Limitations
MemoBrain’s effectiveness is bounded by assumptions about sustained agent reasoning, the limited operations studied, and incomplete baseline coverage.
- MemoBrain assumes the underlying agent can sustain long-horizon tool use and iterative reasoning.Early termination can prevent memory management from triggering and restrict practical benefits.
- The study evaluates only folding and selective flush, leaving richer memory operations and their coordination for future work.Potential extensions include reactivating flushed episodes and partitioning the memory graph for parallel exploration.
- The experimental comparison omits some relevant baselines because certain methods are difficult to adapt, unavailable, or too resource-intensive.Reproduced DeepResearch-30B-A3B results on GAIA and WebWalker are also slightly below original reports because of tool-implementation differences, although comparisons use identical settings.
A.1 Implementation Details
MemoBrain is trained from Qwen3 models with synthesized InfoSeek supervision and evaluated across three benchmarks using specified retrieval, execution, budgeting, and judging configurations.
- GAIA and WebWalker use Google Custom Search, Jina.AI web reading, Qwen3 refinement, and Python exec(), while BrowseComp-Plus uses Qwen3-Embed-8B retrieval.
- Pass@1 is evaluated with GPT-4o-2024-08-06 as the LLM-based judge on experiments run using 8 NVIDIA H100 80G GPUs.
A.2 Prompts
The appendix presents the prompts used for MemoBrain’s memory construction and memory management operations.
- Table 3 presents MemoBrain’s system prompt for executive memory construction.
- Table 4 presents MemoBrain’s system prompt for executive memory management.
A.3 Case Study
The case study shows how MemoBrain constructs and manages a dependency-aware memory graph by compressing reasoning episodes and regulating information exposure.
- Memory graph: A representative graph contains six reasoning episodes rooted at the original input task, with nodes encoding thought types, activation states, compressed representations, and dependencies.
- Case-study outcome: A trajectory spanning tens of thousands of tokens is represented by a few hundred tokens while retaining high-salience conclusions and essential logical dependencies.
- Memory construction: MemoBrain abstracts each episode’s deliberation, tool invocation, and responses into a thought unit before integrating it into the global graph.The process extracts semantic outcomes, assigns thought types and activation states, and establishes dependency relations.
- Memory management: As context approaches its budget, MemoBrain folds completed sub-trajectories into summary thoughts and flushes low-utility or superseded thoughts.These operations regulate which information remains exposed during reasoning.