AMA: Adaptive Memory via Multi-Agent Collaboration

TL;DR

AMA addresses long-term reasoning with adaptive memory by decomposing memory lifecycle into four specialized agents. The Constructor creates multi-granularity memory; the Retriever performs intent-driven routing; the Judge verifies relevance and consistency; the Refresher updates or deletes conflicting memories, maintaining coherence. Empirical results on LoCoMo and LongMemEvals show AMA outperforms baselines and reduces token usage by up to 80%, with strong knowledge-update performance. This work highlights the importance of explicit memory maintenance and adaptive retrieval for scalable, coherent long-horizon reasoning in LLM agents.

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 The Refresher then enforces memory consistency by performing targeted updates or removing outdated entries.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.

Figures (12)

• Figure 1: Comparison of static paradigms and the AMA framework. (a) Static methods suffer from the dilemma of fixed granularity, leading to either noise or information loss. (b) AMA dynamically determines the memory granularity to use, aligning retrieval precision with reasoning demands.
• Figure 2: Overview of the AMA framework. The system orchestrates four agents to enable adaptive memory evolution. The Retriever routes inputs to optimal granularities based on intent. The Judge audits content relevance to trigger feedback loops and detects conflicts. The Refresher executes updates or deletions to rectify these inconsistencies. Finally, the Constructor synthesizes the validated context into structured memory entries.
• Figure 3: Memory Construction Stage. In this stage Constructor generates raw text and fact knowledge memories from utterances, while conditionally synthesizing abstract episodes upon trigger activation.
• Figure 4: Effect of retrieval round limit $K_r$. The left and middle panels show that increasing $K_r$ improves performance on LoCoMo and LongMemEvals with diminishing returns, while the right panel illustrates the corresponding growth in token consumption and inference latency.
• Figure 5: Case Study. (1) The upper part of the figure shows conflict resolution, where outdated factual memories are updated to maintain consistency. (2) The lower part of the figure shows adaptive retrieval, routing queries to different memory types based on intent.