AMA: Adaptive Memory via Multi-Agent Collaboration

2025-07-20 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Adaptive Memory via Multi-Agent Collaboration (AMA) is a novel framework designed to enhance Large Language Model (LLM) agents' long-term memory and reasoning capabilities. It addresses limitations in existing memory systems, such as rigid retrieval granularity and unchecked accumulation of inconsistencies, by employing a multi-agent architecture. AMA features a hierarchical memory design with dynamic retrieval granularity, managed by four coordinated agents: the Constructor, Retriever, Judge, and Refresher. The Constructor builds multi-granular memory (Raw Text, Fact Knowledge, Episode Memory), the Retriever routes queries to appropriate granularities, the Judge verifies relevance and consistency, and the Refresher updates or removes outdated entries. Experiments on LoCoMo and LongMemEvals benchmarks show AMA significantly outperforms state-of-the-art baselines, achieving an LLM Score of 0.774 with GPT-4o-mini and reducing token consumption by approximately 80% compared to full-context methods.

Key takeaway

For AI Architects and NLP Engineers designing advanced LLM agents, AMA offers a robust blueprint for long-term memory management. Its multi-agent, multi-granularity approach significantly improves reasoning accuracy and consistency while drastically cutting token consumption. Consider adopting a similar modular, adaptive memory framework to enhance your agents' performance on complex, long-context tasks and reduce operational costs.

Key insights

AMA uses multi-agent collaboration and hierarchical memory to provide adaptive, consistent long-term memory for LLM agents.

Principles

• Memory granularity should dynamically align with task complexity.
• Explicit conflict detection and resolution are crucial for memory consistency.
• Decomposing memory management into specialized agent roles improves robustness.

Method

AMA orchestrates Constructor, Retriever, Judge, and Refresher agents. The Constructor creates multi-granular memory. The Retriever routes queries. The Judge verifies relevance and consistency. The Refresher updates or deletes conflicting entries.

In practice

• Implement multi-granularity memory (raw, fact, episode) for diverse retrieval needs.
• Use a "Judge" agent to filter irrelevant content and detect logical conflicts.
• Employ a "Refresher" agent for targeted memory updates or deletions.

Topics

• Large Language Model Agents
• Adaptive Memory Systems
• Multi-Agent Collaboration
• Hierarchical Memory Design
• Dynamic Retrieval Granularity

Code references

• langchain-ai/langmem

Best for: AI Architect, NLP Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Multi-Agent Transactive Memory — MATM enables LLM agents to collectively store and retrieve interaction trajectories, improving task performance and efficiency across populations.
• Whitepaper Companion Podcast - Context Engineering: Sessions & Memory — Building adaptive LLM agents requires dynamic context management, session-based conversation history, and long-term memory systems.
• Multi-Agent Memory Is Harder Than You Think — Multi-agent system failures often stem from memory mismanagement, not reasoning flaws, due to confusing retrieval with true memory.
• Agentic Recommender System with Hierarchical Belief-State Memory — Hierarchical memory and adaptive lifecycle management significantly enhance agentic recommender system performance.
• Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity — Memora decouples memory content from retrieval, balancing abstraction and specificity for scalable AI agent memory, achieving leading performance.
• Rosetta Memory: Adaptive Memory for Cross-LLM Agents — Rosetta Memory enables cross-LLM agents to share and adapt memory effectively, shifting from LLM-centric to memory-centric design.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.