AMA: Adaptive Memory via Multi-Agent Collaboration

Abstract: The rapid evolution of 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 The Refresher then enforces memory consistency by performing targeted updates or removing outdated entries.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.

• The paper introduces the AMA framework, a multi-agent system for managing external long-term memory in LLM agents, achieving state-of-the-art results on LoCoMo and LongMemEval while using 80% fewer tokens than full-context inference.
• AMA decomposes the memory lifecycle into four coordinated roles—Constructor, Retriever, Judge, and Refresher—that manage three different granularities and dynamically adapt retrieval and update tasks.
• AMA’s efficiency stems from its dynamic knowledge maintenance and conflict resolution, highlighting its 33% improvement in knowledge-update accuracy over baselines when including the Refresher component.

Overview

AMA (Adaptive Memory via Multi-Agent Collaboration) is a framework for external long-term memory in LLM agents that addresses two shortcomings of prior agentic memory systems: a mismatch between fixed storage/retrieval granularity and task-specific reasoning demands, and the unchecked accumulation of logical inconsistencies under accumulation-heavy maintenance strategies (2601.20352). The paper decomposes the memory lifecycle into four coordinated agent roles—Constructor, Retriever, Judge, and Refresher—and demonstrates state-of-the-art results on LoCoMo and LongMemEval while consuming roughly 80% fewer tokens than full-context inference.

Motivation and positioning

Existing external memory systems rely on static text chunking or coarse summaries, which either inject irrelevant noise (overly coarse retrieval) or fragment logical dependencies across isolated chunks (excessively fine granularity). Agentic memory frameworks such as MemGPT, Mem0, and A-Mem improve representational flexibility but lack an explicit adaptive routing mechanism at inference time and depend on coarse-grained updates that allow redundancy and errors to accumulate. The closest related system, MIRIX, also assigns specialized agents for memory organization, but the authors exclude it as a baseline because its official implementation was unavailable during their experiments; they note it lacks dedicated mechanisms for long-term memory consistency. AMA's distinguishing claim is that fine-grained control over retrieval, verification, and memory evolution is difficult to achieve within a monolithic controller without entangling conflicting objectives.

Framework design

The Constructor transforms unstructured dialogue into three granularities. Raw Text Memory preserves utterances verbatim with reference indices and metadata. Fact Knowledge Memory decomposes input into atomic facts constrained to five SVO-based sentence patterns (S-V, S-V-O, S-V-C, S-V-O-O, S-V-O-C), each stored as an independent unit with source-turn references and timestamps. Episode Memory is synthesized conditionally via a binary trigger activated by topic shift, explicit user request, or context-window saturation. All entries are embedded with text-embedding-3-large for similarity-based retrieval.

The Retriever rewrites queries into self-contained forms and emits a four-dimensional binary intent vector encoding fine-grained detail, abstract summary, cross-temporal event, and atomic-fact requirements. A priority-based routing function maps this vector to a repository: fine-grained intents route to Raw Text, abstract or event-level intents to Episodes, and everything else defaults to Fact Knowledge. The retrieval budget $K_{dyn}$ is predicted dynamically subject to a minimum threshold.

The Judge performs dual verification on retrieved candidates. Relevance assessment triggers a bounded Retry loop (up to $K_r$ rounds) that traverses alternative granularities or expands retrieval via relation indices when information density is insufficient. Conflict detection identifies contradictions between the current input and filtered memories—for example, outdated user-state facts—and isolates a conflict set that activates the Refresher.

The Refresher executes targeted maintenance with strict conditional branching: Delete is permitted only under explicit user instruction or exceeded retention lifespan; all other conflicts default to attribute-level Update operations that preserve memory continuity. This conservative policy is central to the framework's consistency guarantees.

Empirical results

On LoCoMo with GPT-4o-mini, AMA achieves an overall LLM Score of 0.774 versus 0.740 for the strongest baseline Nemori; with GPT-4.1-mini it reaches 0.805, making it the only evaluated method to surpass FullContext (0.786). With Qwen3-30B-Instruct, AMA attains 0.791 against FullContext's 0.733, and even the smaller Qwen3-8B backbone yields 0.707 versus 0.696 for FullContext. On LongMemEval$_s$, AMA reaches 0.698 average accuracy, exceeding Nemori by 0.056 and Zep by 0.066, with near-perfect performance on single-session-user questions (0.986).

Method	Tokens	Latency (s)	LLM Score
FullContext	18625	7.21	0.717
RAG	5800	2.98	0.300
Nemori	2925	3.15	0.740
Mem0	1340	3.74	0.608
AMA ($K_r{=}1$)	2491	3.12	0.723
AMA ($K_r{=}2$)	3613	3.91	0.774

At $K_r = 2$, AMA uses approximately 19% of FullContext's tokens while achieving the highest score among memory frameworks—an efficiency–accuracy trade-off the authors identify as the practical operating point. Performance saturates beyond $K_r \geq 5$ while token cost and latency grow roughly linearly.

Ablations isolate two findings. First, multi-granularity storage is complementary: Fact Knowledge alone is the strongest single granularity (0.712 LLM Score on LoCoMo), but the full three-granularity configuration outperforms any subset. Second, the Refresher is decisive for dynamic knowledge maintenance: removing it drops knowledge-update accuracy on LongMemEval from 0.897 to 0.568—a 33-point degradation that directly supports the paper's argument that conflict resolution, not merely richer storage, is required for reliable long-term memory.

Limitations

The authors concede that multi-agent collaboration incurs moderate computational overhead relative to static retrieval baselines, and that performance depends on the backbone model's reasoning capability, leaving efficiency on smaller architectures as an open optimization problem. Two further caveats bear on interpretation: MIRIX was not compared due to implementation unavailability, so the claimed advantage over specialized multi-agent memory systems rests on qualitative rather than empirical grounds; and LongMemEval evaluation was restricted to GPT-4o-mini for computational feasibility, limiting evidence of generalization on that benchmark to a single backbone. The Judge's conflict detection is prompt-driven and bounded by one-step inference, so its recall on subtle or distributed contradictions is not separately quantified.

Conclusion

AMA contributes a multi-agent decomposition of the memory lifecycle—multi-granularity construction, intent-based adaptive routing, verified retrieval with feedback loops, and conservative conflict-driven maintenance—that yields consistent gains over strong baselines across four backbones and two benchmarks, with substantial token savings over full-context inference. The ablation evidence indicates that explicit consistency mechanisms account for much of the advantage in knowledge-update scenarios. Open questions include reducing multi-agent overhead for small backbones and empirically validating the framework against other specialized multi-agent memory systems.