AMA: Adaptive Memory via Multi-Agent Collaboration
AMA framework achieves adaptive memory via multi-agent collaboration, reducing token consumption by 80%.
Key Findings
Methodology
AMA framework uses multi-agent collaboration to manage memory, including Constructor, Retriever, Judge, and Refresher. Constructor builds multi-granularity memory, Retriever dynamically routes queries, Judge verifies relevance and consistency, Refresher updates or removes outdated entries.
Key Results
- AMA significantly outperforms state-of-the-art baselines in long-context benchmarks, reducing token consumption by approximately 80% while maintaining high precision and long-term consistency.
- In knowledge update scenarios, AMA achieves nearly 90% accuracy, significantly outperforming existing methods.
- By adaptively controlling retrieval granularity, AMA excels across multiple scenarios.
Significance
This study addresses the mismatch of granularity and logical inconsistencies in existing memory systems through AMA framework, advancing the application of LLMs in complex reasoning and long-term interaction.
Technical Contribution
AMA achieves fine-grained control over the memory lifecycle through multi-agent collaboration, offering higher flexibility and precision compared to existing single-controller methods.
Novelty
AMA is the first to achieve adaptive memory granularity control through multi-agent collaboration, addressing inherent issues of static storage granularity.
Limitations
- In extremely complex tasks, AMA may require more computational resources to maintain efficient memory management.
- In certain cases, the Judge may not fully detect all logical conflicts.
Future Work
Future research could explore AMA's performance in broader application scenarios and optimize its computational efficiency.
AI Executive Summary
The rapid evolution of Large Language Models (LLMs) necessitates robust memory systems to support long-term interaction and complex reasoning. Existing methods often rely on fixed retrieval granularity and coarse-grained update mechanisms, leading to a mismatch between stored information and task demands. To address these issues, researchers propose the AMA framework, which achieves memory management through multi-agent collaboration. The framework employs a hierarchical memory design, dynamically adjusting retrieval granularity to match task complexity. Experimental results show that AMA significantly outperforms state-of-the-art baselines in long-context benchmarks while reducing token consumption by approximately 80%. This study not only enhances retrieval precision but also ensures long-term memory consistency, laying a solid foundation for the continued evolution of LLMs. Future research can further optimize AMA's computational efficiency and explore its potential in more application scenarios.
Deep Analysis
Background
As Large Language Models (LLMs) demonstrate strong capabilities in complex reasoning and multi-turn interaction scenarios, long-term memory systems become crucial to support these behaviors. Existing memory approaches are broadly categorized into internal and external paradigms. Internal memory implicitly absorbs historical information into model parameters but is constrained by capacity and update costs. External memory relies on explicit storage and retrieval, offering superior scalability and editability, becoming the dominant approach.
Core Problem
Existing memory systems typically rely on fixed retrieval granularity and coarse-grained update mechanisms, leading to a mismatch between stored information and task-specific reasoning demands, while accumulating logical inconsistencies over time.
Innovation
The AMA framework achieves memory management through multi-agent collaboration, employing a hierarchical memory design that dynamically adjusts retrieval granularity to match task complexity. It offers higher flexibility and precision compared to existing single-controller methods.
Methodology
- �� Constructor: Generates multi-granularity memory including Raw Text, Fact Knowledge, and Episode Memory.
- �� Retriever: Dynamically routes queries to the most appropriate memory form.
- �� Judge: Verifies the relevance and consistency of retrieved content.
- �� Refresher: Performs targeted updates or removes outdated entries.
Experiments
Experiments were conducted using LoCoMo and LongMemEvals benchmarks, comparing AMA with various baseline methods. Metrics such as F1 and BLEU-1 were used to evaluate memory retrieval precision and consistency.
Results
AMA significantly outperforms state-of-the-art baselines in long-context benchmarks, reducing token consumption by approximately 80% while maintaining high precision and long-term consistency. In knowledge update scenarios, AMA achieves nearly 90% accuracy.
Applications
The AMA framework can be applied in scenarios requiring long-term memory and complex reasoning, such as intelligent assistants and automated customer service systems.
Limitations & Outlook
In extremely complex tasks, AMA may require more computational resources to maintain efficient memory management. In certain cases, the Judge may not fully detect all logical conflicts.
Plain Language Accessible to non-experts
Imagine a library with many bookshelves, each holding different types of books. AMA is like a smart librarian who can quickly find the most suitable book based on the reader's needs and ensure the book information is up-to-date. The Constructor is responsible for placing new books on the right shelves, the Retriever finds the books, the Judge ensures the information is accurate, and the Refresher updates or removes outdated books. This way, the library's books are always well-organized, and readers can always find the information they need.
ELI14 Explained like you're 14
Imagine you're playing a super complex game where you need to remember lots of things to win. AMA is like your game assistant, helping you remember all the important details. The Constructor organizes all the information, the Retriever helps you find what you need, the Judge makes sure the info is correct, and the Refresher updates outdated info. This way, you can focus on the game without worrying about forgetting important stuff!
Glossary
Constructor
Responsible for generating multi-granularity memory including Raw Text, Fact Knowledge, and Episode Memory.
Used in the memory construction stage to ensure structured storage of information.
Retriever
Dynamically routes queries to the most appropriate memory form, ensuring retrieval precision.
Used in the memory access stage to optimize query relevance.
Judge
Verifies the relevance and consistency of retrieved content, triggering feedback loops or refresh operations.
Used to ensure logical consistency of memory.
Refresher
Performs targeted updates or removes outdated entries to maintain memory consistency.
Used when logical conflicts are detected to ensure information accuracy.
Multi-Agent System
Achieves complex task collaboration and decision-making through role specialization.
Used in AMA framework for memory management.
Open Questions Unanswered questions from this research
- 1 How to optimize AMA's computational efficiency in extremely complex tasks?
- 2 In certain cases, the Judge may not fully detect all logical conflicts; how can this be improved?
Applications
Immediate Applications
Intelligent Assistant
AMA can be used in intelligent assistants, providing more accurate long-term memory and complex reasoning support.
Long-term Vision
Automated Customer Service
Through AMA framework, customer service systems can achieve more efficient memory management and information updates.
Abstract
The rapid evolution of Large Language Model (LLM) agents has necessitated robust memory systems to support cohesive long-term interaction and complex reasoning. Benefiting from the strong capabilities of LLMs, recent research focus has shifted from simple context extension to the development of dedicated agentic memory systems. However, existing approaches typically rely on rigid retrieval granularity, accumulation-heavy maintenance strategies, and coarse-grained update mechanisms. These design choices create a persistent mismatch between stored information and task-specific reasoning demands, while leading to the unchecked accumulation of logical inconsistencies over time. To address these challenges, we propose Adaptive Memory via Multi-Agent Collaboration (AMA), a novel framework that leverages coordinated agents to manage memory across multiple granularities. AMA employs a hierarchical memory design that dynamically aligns retrieval granularity with task complexity. Specifically, the Constructor and Retriever jointly enable multi-granularity memory construction and adaptive query routing. The Judge verifies the relevance and consistency of retrieved content, triggering iterative retrieval when evidence is insufficient or invoking the Refresher upon detecting logical conflicts. The Refresher then enforces memory consistency by performing targeted updates or removing outdated entries. Extensive experiments on challenging long-context benchmarks show that AMA significantly outperforms state-of-the-art baselines while reducing token consumption by approximately 80% compared to full-context methods, demonstrating its effectiveness in maintaining retrieval precision and long-term memory consistency.