HMARS: A Hierarchical Multi-Agent Memory System for Long-Context Reasoning
HMARS enhances long-context reasoning by using a hierarchical multi-agent memory system, significantly improving evidence coverage.
Key Findings
Methodology
HMARS addresses long-context reasoning by organizing inputs as a hierarchical multi-agent memory system. Sub-agents manage local memory access, mid-agents coordinate regional context and query-specific guidance, and a frontier model performs final reasoning over retrieved evidence pages. This method prevents premature evidence discarding seen in traditional retrieval approaches.
Key Results
- HMARS achieved the best performance in long-document and multi-turn dialogue tasks, with an evidence coverage rate of 99.4%, significantly outperforming traditional RAG methods.
- In multi-turn memory tasks, HMARS achieved an accuracy of 90.8%, showing significant improvement over other baseline methods.
- Ablation studies indicate that HMARS's advantage primarily stems from improved evidence access rather than changes in final answer prompting.
Significance
HMARS offers a new perspective on long-context reasoning by treating long documents and interaction histories as managed memory systems rather than flat retrieval corpora. This approach holds significant implications for academia and industry, especially in complex tasks requiring the integration of scattered evidence.
Technical Contribution
By introducing a hierarchical multi-agent architecture, HMARS provides a distinct technical path from existing methods. It not only improves evidence coverage but also introduces new engineering possibilities through regional context management and query coordination.
Novelty
HMARS is the first to treat long contexts as managed memory systems rather than simple retrieval problems. Its core innovation lies in its hierarchical structure and multi-agent collaboration, setting it apart from existing methods.
Limitations
- In certain reasoning types, HMARS underperforms compared to high-budget RAG settings, particularly in inference tasks.
- The system's complexity may lead to increased computational costs.
- Further research is needed to explore its applicability across different domains.
Future Work
Future work could explore HMARS's application in more domains and optimize its computational efficiency. Additionally, research could focus on enhancing the system's flexibility and adaptability.
AI Executive Summary
Long-context reasoning requires models to access, retrieve, and integrate evidence scattered across documents, dialogues, and interaction histories. Existing retrieval-augmented generation methods often reduce this to top-K chunk retrieval, but this passive access can discard relevant evidence before reasoning begins. HMARS proposes a hierarchical multi-agent memory system that treats long contexts as managed memory rather than a flat retrieval corpus. Sub-agents maintain grounded access to bounded memory regions, mid-agents manage regional context and provide query-specific coordination, and a frontier model performs final reasoning over retrieved evidence pages. Experimental results show that HMARS achieves the best performance in long-document and multi-turn memory tasks, significantly improving evidence coverage. Its advantage primarily comes from retrieving the required supporting evidence more completely rather than merely changing the final answer prompt. Although HMARS underperforms in certain reasoning types compared to high-budget RAG settings, it offers a new solution for complex tasks requiring the integration of scattered evidence. Future work could explore its application in more domains and optimize computational efficiency.
Deep Analysis
Background
Long-context reasoning is crucial in natural language processing, especially for complex tasks requiring the integration of scattered evidence. Traditional methods like retrieval-augmented generation (RAG) often rely on top-K chunk retrieval, which has limitations in handling long documents and multi-turn dialogues, as important evidence can be discarded.
Core Problem
The core problem in long-context reasoning is effectively accessing and integrating evidence scattered across different document regions. Traditional methods struggle with long documents, often discarding important information and leading to inaccurate reasoning.
Innovation
HMARS's core innovation is its hierarchical multi-agent memory system. By treating long documents as managed memory systems, HMARS avoids the premature evidence discarding seen in traditional methods. Sub-agents manage local memory access, mid-agents coordinate regional context and provide query-specific guidance.
Methodology
- �� Sub-agents manage local memory access, ensuring every region participates in query assessment.
- �� Mid-agents manage regional context, providing query coordination to help retrieve focused evidence.
- �� The frontier model performs final reasoning over retrieved evidence pages, ensuring reasoning accuracy.
Experiments
Experiments were conducted on long-document and multi-turn dialogue tasks using LongBench-v2 and RealMem datasets. Baseline methods included RAG, Dense Rerank, and GraphRAG. Evaluation metrics included evidence coverage and reasoning accuracy.
Results
HMARS achieved an evidence coverage rate of 99.4% in long-document tasks and an accuracy of 90.8% in multi-turn memory tasks. Compared to other baseline methods, HMARS performed best in most reasoning types, especially in tasks requiring the integration of scattered evidence.
Applications
HMARS is suitable for complex tasks requiring the integration of scattered evidence, such as legal document analysis and scientific literature reviews. Its hierarchical structure and multi-agent collaboration provide significant advantages in these fields.
Limitations & Outlook
While HMARS excels in most tasks, it underperforms in certain reasoning types compared to high-budget RAG settings. Additionally, its system complexity may lead to increased computational costs. Future work could explore its application in more domains and optimize computational efficiency.
Plain Language Accessible to non-experts
Imagine you're in a large library looking for specific information. Traditional methods are like randomly picking a few books off the shelf, hoping to find the answer. HMARS, however, is like an organized team where each member is responsible for a specific section and provides relevant books based on your question. This ensures you don't miss any important information and find the answer faster.
ELI14 Explained like you're 14
Imagine playing a massive multiplayer online game where you need to find clues scattered across the map to complete a mission. Traditional methods are like running around the map randomly, hoping to find all the clues. HMARS is like having a team where each member is responsible for an area, helping you find all the clues you need to complete the mission. Cool, right?
Glossary
HMARS (Hierarchical Multi-Agent Memory System)
A system for long-context reasoning that manages and retrieves evidence through a hierarchical multi-agent structure.
Used to enhance evidence coverage in long-document and multi-turn dialogue tasks.
RAG (Retrieval-Augmented Generation)
A technique that enhances generative models by retrieving relevant text chunks.
Used as a baseline method for comparison with HMARS.
Sub-Agent
An agent in HMARS responsible for accessing local memory regions.
Ensures each region participates in query assessment.
Mid-Agent
An agent in HMARS that manages regional context and provides query coordination.
Helps retrieve focused evidence.
Frontier Model
The model in HMARS that performs final reasoning.
Conducts reasoning over retrieved evidence pages.
Open Questions Unanswered questions from this research
- 1 How can HMARS's computational efficiency be further optimized? The current system complexity is high, leading to significant computational costs.
- 2 What is HMARS's applicability across different domains? More experiments are needed to verify its effectiveness.
Applications
Immediate Applications
Legal Document Analysis
HMARS can be used to analyze complex legal documents, helping lawyers quickly find relevant evidence.
Long-term Vision
Scientific Literature Review
HMARS can integrate scattered evidence in scientific literature, aiding researchers in comprehensive reviews.
Abstract
Long-context reasoning requires models to access, retrieve, and integrate evidence scattered across documents, dialogues, and accumulated interaction histories. Standard retrieval-augmented generation reduces this problem to top-$K$ chunk retrieval, but such passive access can discard relevant evidence before reasoning begins, especially when relevance depends on broader context. We propose HMARS, a hierarchical multi-agent memory system that treats long contexts as managed memory rather than a flat retrieval corpus. Sub-agents maintain grounded access to bounded memory regions, mid-agents manage regional context and provide query-specific coordination, and a frontier model performs final reasoning over retrieved evidence pages. To evaluate this view, we construct two diagnostic benchmarks targeting evidence breadth and context-dependent relevance. Across long-document and multi-turn memory tasks, HMARS achieves the best overall performance against retrieval, reranking, full-context, graph-based, and agentic long-context baselines. Evidence coverage analysis further shows that its gains come from retrieving the required supporting evidence more completely, rather than merely changing the final answer prompt.