🤖 AI

HyMem: Hierarchical Context Management for Long-Horizon Agents via Information Isolation

HyMem is a hierarchical context management framework that improves long-horizon LLM agent performance by isolating high-level planning from execution traces and using structured summaries to maintain focus and accuracy, achieving state-of-the-art results on GAIA and Browsecomp-plus benchmarks.

Original authors: XinQi Wang, Jinwei Xiao, Sijia Cui, Hongming Zhang, Yanna Wang, Qingyang Zhang, Bo Xu

Published 2026-08-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: XinQi Wang, Jinwei Xiao, Sijia Cui, Hongming Zhang, Yanna Wang, Qingyang Zhang, Bo Xu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a brilliant assistant who can read, reason, and search the internet to solve incredibly difficult problems. For years, scientists have been building these digital helpers using large language models, hoping they could tackle complex, multi-step challenges like planning a scientific expedition or navigating a website to find a specific historical fact. However, a persistent problem has held them back: as these assistants work longer and gather more information, their "memory" of the conversation becomes a cluttered mess. The assistant starts to drown in the details of its own actions—the raw text of every search result, every failed attempt, and every intermediate thought—until it forgets the original goal it was trying to reach. It is as if a librarian, while trying to find a single book, gets buried under a mountain of every page they have ever turned, losing track of the title they were looking for.

Researchers at the Institute of Automation in the Chinese Academy of Sciences have proposed a new way to organize this digital memory to solve that problem. They call their system HyMem, a method designed to keep long-term tasks on track by strictly separating different types of information. Instead of letting every piece of data mix together in one long list, HyMem builds a structure where high-level planning happens in a clean, quiet room, while the noisy, messy work of searching and testing happens in separate, isolated spaces. This approach allows the assistant to focus on the big picture without getting distracted by the clutter of its own daily operations.

The core idea behind this work is that not all information is created equal. When an agent works on a long task, it generates two very different kinds of data. One type is the strategic plan: the goals, the verified facts, and the steps needed to finish the job. The other type is the execution trace: the raw, often repetitive, and sometimes failed attempts to use tools, browse websites, or check details. In older systems, these two types of information were mixed together in a single stream. As the task went on, the sheer volume of the messy execution traces would overwhelm the sparse, critical planning signals. The assistant would lose its way, unable to distinguish the important goal from the noise of its own history.

To fix this, the researchers designed a framework that acts like a strict gatekeeper for information. They created a "Planner" layer that holds the main strategy and a separate "Executor" layer that handles the actual work of using tools. When the Planner needs to check a fact or run a search, it sends a specific instruction to the Executor. The Executor goes off and does the work, gathering all the raw data, but it does not dump that raw data back into the Planner's main memory. Instead, the Executor processes the information, filters out the noise, and returns only a clean, structured summary of what it found. This summary is then added to the Planner's memory, while the messy details are discarded.

The system also includes a special "Isolated Reasoning" module for particularly tricky sub-problems. If a task requires deep thinking or checking conflicting evidence, the Planner sends that specific problem to this isolated module. The module works through the problem on its own, keeping its own internal thoughts and intermediate steps hidden from the main Planner. Once it reaches a conclusion, it sends back just the answer and the key evidence, leaving the main Planner's memory free of the long, winding path taken to get there. This ensures that the main decision-maker never gets bogged down by the complexity of the sub-tasks it delegated.

To keep the system running smoothly over long periods, HyMem also uses a structured memory system that acts like a filing cabinet. It organizes past events, current goals, and lessons learned about how to use tools into neat, compact summaries. When the system needs to remember what it has done, it consults these organized files rather than trying to recall every single detail of the past. This allows the assistant to maintain a clear sense of progress and continuity, even after many hours of work, without its memory becoming too large to manage.

The researchers tested this new approach on two challenging sets of tasks: one involving general reasoning and tool use, and another focused on deep research questions that required finding specific facts across the internet. They compared their system against several other methods, including older approaches that simply tried to compress the entire history of a conversation into a shorter summary. The results showed that by keeping the planning and execution separate, the HyMem system was significantly more successful. On the general reasoning tasks, it solved about 66.7% of the problems correctly, while on the deep research tasks, it solved 61.3%. These numbers were notably higher than the best-performing alternative methods, which struggled to maintain focus as the tasks grew longer and more complex.

The study also looked at how much information the system had to process to reach these answers. The researchers found that the HyMem system did not simply solve problems by using more computing power or reading more text. Instead, it solved them more efficiently by directing its attention only to the information that mattered. The main planning memory grew very slowly, staying focused on the goal, while the isolated spaces handled the heavy lifting of exploration and analysis. This suggests that the key to solving long, difficult problems is not just having a bigger memory, but having a smarter way to organize it.

While the results are promising, the researchers note that the system relies on the underlying artificial intelligence being able to follow strict instructions and produce well-organized summaries. If the model cannot follow these rules, the system's benefits may not appear. However, for tasks that require sustained attention and the ability to navigate complex information over time, this hierarchical approach offers a clear path forward. By separating the signal from the noise, HyMem allows digital agents to think clearly, even when the work they are doing is incredibly complicated.

Technical Summary: HyMem – Hierarchical Context Management for Long-Horizon Agents

1. Problem Statement

Large Language Model (LLM) agents struggle with complex, long-horizon tasks due to context dilution. As interactions accumulate, the context window becomes dominated by dense, noisy execution traces (e.g., raw tool outputs, browsing logs, intermediate deliberations) and transient details. This "information flood" obscures sparse but critical high-level planning signals (goals, constraints, verified facts), causing the agent to lose track of its strategic objectives.

Existing solutions, such as context compression or external memory retrieval, typically operate on a single, flat sequence of interactions. These methods fail to fundamentally separate strategic planning from execution noise, often leading to the loss of critical reasoning chains or the inability to distinguish between high-level goals and low-level tool usage. The core issue is not merely context length, but the information-density mismatch caused by mixing heterogeneous information streams in a single sequence.

2. Methodology: The HyMem Framework

To address this, the authors propose HyMem, a training-free, inference-time framework that enforces hierarchical context isolation. Instead of treating context as a passive log, HyMem actively manages information flow by decoupling the agent's cognitive processes into distinct functional layers and context spaces.

Core Architecture

HyMem separates the agent's operation into four distinct components, ensuring that raw execution traces and intermediate reasoning do not directly pollute the main planning context:

Planner Layer: Maintains the persistent, high-level decision context. It holds only distilled, schema-constrained returns from other modules and structured memory summaries. It is responsible for task decomposition, strategic planning, and synthesizing results.
Executor Layer: Handles tool interactions and low-level execution. Upon receiving a directive, it creates a fresh, isolated context to perform tool calls. Raw observations are distilled (removing navigation text, errors, and irrelevant snippets) before being synthesized into a structured RESULTS message. Only this distilled result is returned to the Planner; the raw context is discarded.
Isolated Reasoning Module: Designed for complex subtasks requiring multi-hop analysis, hypothesis checking, or evidence reconciliation. When triggered, it spawns a separate reasoning session with its own context. The intermediate deliberation traces remain private to this session. Only a structured RETURN message containing conclusions, supporting facts, and confidence scores is passed back to the Planner.
Memory Management Module: Compresses historical trajectories into three rewritable memory structures to preserve continuity across context refreshes:
- Episode Memory: Records completed milestones, verified facts, and major decisions.
- Working Memory: Tracks current objectives, unresolved gaps, and immediate next steps.
- Tool Memory: Stores effective tool-use patterns, failed queries, and domain-specific caveats.

Operational Mechanism

The framework operates via typed context isolation. The Planner interacts with the Executor and Isolated Reasoning modules only through schema-constrained messages (e.g., <TASK>, <ISOREASON>, <RESULTS>, <RETURN>).

Context Boundaries: Raw observations and intermediate reasoning tokens are strictly confined to their respective temporary contexts (Executor or Isolated Reasoning). They are deleted once a structured return is generated.
Memory Folding: A FOLD operator is triggered when specific conditions are met (e.g., high token utilization, consecutive failures, or turn limits). This rewrites the memories, ensuring the Planner context grows slowly and remains focused on task-relevant information rather than accumulating a linear history of all interactions.

3. Key Contributions

The paper formalizes the problem of long-horizon reasoning as an information-density mismatch and introduces the following contributions:

Formalization of Context Dilution: The authors define the degradation of agent performance as a reduction in the relative density of planning-relevant information ( $\rho_t$ ) caused by the rapid accumulation of execution traces ( $|X_t|$ ).
HyMem Framework: A training-free, plug-and-play architecture that implements hierarchical context isolation. It explicitly separates the Planner, Executor, Isolated Reasoning, and Memory Management into distinct spaces, preventing the propagation of execution noise into the decision-making space.
Empirical Validation: Comprehensive evaluations on the GAIA and Browsecomp-plus benchmarks demonstrate that HyMem significantly outperforms strong baselines (ReAct, ReSum, A-MEM, THREAD) while maintaining comparable resource consumption.

4. Experimental Results

The framework was evaluated using DeepSeek-V4 and Qwen3-32B backbones on two datasets:

GAIA (General AI Assistants): HyMem with DeepSeek-V4 achieved an average Pass@1 of 66.7%, outperforming the ReAct baseline (56.6%) by 10.1 percentage points and the ReSum baseline by 6.1 percentage points. The improvement was consistent across all difficulty levels, with the largest gains on Level 3 (hard) questions.
Browsecomp-plus (Deep Research): HyMem with DeepSeek-V4 achieved an average Pass@1 of 61.3%, surpassing ReAct (56.6%) by 4.7 percentage points. Notably, on hard tasks, Pass@1 increased from 14.0% to 30.0%.
Ablation Studies: Removing the Isolated Reasoning module caused a significant performance drop (5.0 pp on Browsecomp-plus, 18.6 pp on GAIA), confirming that isolating complex subtask analysis is crucial. Similarly, removing Memory Management degraded performance, highlighting the need for structured continuity.
Efficiency: Analysis of context growth (Figure 4) reveals that while the total context (including temporary modules) grows, the Planner's persistent context grows much more slowly than in flat architectures like ReAct. This allows for extensive exploration without diluting the strategic focus.

5. Significance and Claims

The paper claims that effective long-horizon reasoning requires selective information isolation combined with continuity preservation, rather than simple context compression.

Architectural Shift: HyMem shifts the paradigm from managing a single, mixed trajectory to managing multiple, isolated information streams. This fundamentally blocks the "information flood" that submerges strategic signals.
Training-Free Advantage: As a training-free solution, HyMem offers immediate applicability to existing LLMs without the need for fine-tuning or learning complex memory operations.
Scalability: The framework demonstrates that agents can maintain focus and accuracy over long, complex tasks by keeping the decision context clean, allowing for sustained reasoning clarity even as the total number of interactions increases.

The authors note that HyMem's effectiveness depends on the backbone model's ability to follow typed control instructions and produce reliable structured returns. While it shows robust gains with stronger models like DeepSeek-V4, performance on smaller models (e.g., Qwen3-32B) on certain benchmarks (Browsecomp-plus) was more nuanced, suggesting that the benefits of context isolation are maximized when paired with models capable of rigorous instruction following.