Findings of ACL 2026Official Code

Memory that adapts
as dialogue evolves.

AMA coordinates four specialized agents to construct, retrieve, verify, and refresh long-term memory—aligning the granularity of evidence with the reasoning task at hand.

Read the paper ↗ Explore the code ↗ arXiv:2601.20352

0.774LoCoMo LLM score

0.698LongMemEval average

81%fewer input tokens*

Research reference code

The paper implementation, evaluation entry points, figures, and reproduction guidance for studying AMA.

View research code ↗

Memory skill & plugin

An AMA-based memory layer for OpenClaw and Harness Agent-style architectures, with automatic recall and capture.

See the integration ↓

01 / The challenge

Long-term memory is not just a retrieval problem.

Fine details, stable facts, and high-level episodes demand different representations. As user information changes, a useful system must detect conflicts and repair stale knowledge—not merely append more context.

System overviewRetriever → Judge → Refresher → Constructor

Overview of the AMA multi-agent memory framework

02 / Collaboration

Retriever

Rewrites ambiguous queries and routes each intent to raw text, fact knowledge, or episode memory.

Judge

Audits relevance and consistency, triggering bounded retrieval feedback when evidence is insufficient.

Refresher

Resolves logical conflicts through targeted updates or deletions, preserving knowledge over time.

Constructor

Transforms validated dialogue into traceable raw turns, atomic facts, and event-level episodes.

Memory construction

Three views of the same evolving conversation.

The Constructor preserves exact turns as raw text, distills reusable facts, and synthesizes cross-turn episodes. Episodes are created when the topic shifts, the user explicitly asks to consolidate, or context becomes saturated.

Retrieve Rewrite the query, infer its intent, and select a memory granularity.
Judge Check relevance and consistency; retry when the evidence is insufficient.
Refresh Update or delete stale memories when a conflict is detected.
Construct Store the validated interaction in complementary forms.

Constructor detailRaw → Fact → Episode

How AMA constructs raw-text, fact-knowledge, and episode memories

Three complementary views

One memory does not fit every question.

AMA stores dialogue at multiple levels and selects a representation according to current reasoning intent.

Raw text precise wording and temporal details
Fact knowledge stable, atomic information
Episode memory events and cross-turn abstraction

Comparison of raw text, fact knowledge, and episode memory

03 / Main results

Higher accuracy,
less context.

LoCoMo · GPT-4o-mini0.774

Overall LLM score, compared with 0.740 for the strongest listed memory baseline.

LongMemEval_s0.698

Average accuracy across six question types, compared with 0.642 for Nemori.

Knowledge update0.897

Accuracy on questions that require the system to follow evolving user information.

Default K_r = 23,613

Input tokens versus 18,625 for FullContext in the efficiency analysis.

*3,613 tokens is approximately 19% of the 18,625-token FullContext input reported in the paper.

04 / Benchmarks

Consistent gains across model families.

AMA is evaluated on LoCoMo (1,540 questions over long multi-session conversations) and LongMemEval_s (500 questions with approximately 115K-token histories). Temperature is 0, retrieval uses top-k 10, and AMA defaults to at most two retrieval rounds.

LoCoMoOverall LLM Score

Backbone	FullContext	Nemori	AMA
GPT-4o-mini	0.717	0.740	0.774
GPT-4.1-mini	0.786	0.774	0.805
Qwen3-30B-Instruct	0.733	0.756	0.791
Qwen3-8B-Instruct	0.696	0.686	0.707

LongMemEval_sAccuracy by question type

Question type	FullContext	Zep	Nemori	AMA
Single-session preference	0.300	0.533	0.467	0.467
Single-session assistant	0.818	0.750	0.839	0.964
Temporal reasoning	0.365	0.541	0.617	0.444
Multi-session	0.406	0.474	0.511	0.624
Knowledge update	0.769	0.744	0.615	0.897
Single-session user	0.814	0.929	0.886	0.986
Average	0.548	0.632	0.642	0.698

AMA leads the average and excels on assistant-side, multi-session, knowledge-update, and user-specific questions. Temporal reasoning remains an important opportunity for future work.

05 / Analysis

A bounded loop finds the useful evidence.

Most of the performance gain arrives in the first few retrieval rounds and then saturates, while token use and latency continue to grow approximately linearly. AMA therefore uses K_r = 2 as its default balance.

LoCoMo performance across retrieval rounds — LoCoMo performance

LongMemEval performance across retrieval rounds — LongMemEval_s performance

Token and latency costs across retrieval rounds — Retrieval cost

Accuracy–cost trade-offLoCoMo · GPT-4o-mini

Method	Tokens	Latency	Score
FullContext	18,625	7.206 s	0.717
RAG	5,800	2.983 s	0.300
Nemori	2,925	3.152 s	0.740
AMA (K_r=1)	2,491	3.124 s	0.723
AMA (K_r=2)	3,613	3.910 s	0.774

What each component addsSelected ablation results

Memory setting	LoCoMo	Knowledge update
Raw text only	0.669	0.767
Fact knowledge only	0.712	0.804
Episode memory only	0.688	0.748
All granularities, no Refresher	0.771	0.568
Full AMA	0.774	0.897

Multi-granular memory gives broad coverage; the Refresher is especially important when user information changes.

06 / Case study

Remember the latest truth—and the right level of detail.

Conflict-aware refresh. When device information changes, the Retriever recalls the earlier fact, the Judge identifies the conflict, and the Refresher updates the stale memory instead of keeping contradictory records.

Intent-aware granularity. A factual query routes to Fact Knowledge; an abstract summary query retrieves Episode Memory.

AMA case study showing conflict-aware refresh and intent-aware memory retrieval — Examples from the paper. The system couples retrieval with explicit relevance judgment and memory maintenance.

Beyond the benchmark

AMA, packaged as a memory skill.

The same repository also turns the research design into a usable memory layer for OpenClaw and Harness Agent-style systems: an agent-facing skill, automatic recall and capture hooks, six explicit tools, and a local Python sidecar backed by SQLite and FAISS.

Two distinct deliverables: the Python research code serves as a reference implementation of the paper; the TypeScript plugin and skill package the design for agent runtimes. They share the AMA concepts but can be explored independently.

shell

$ git clone https://github.com/Sherlockwz/AMA.git
$ cd AMA && ./scripts/setup.sh
$ openclaw plugins install --link .
$ openclaw ama doctor

OpenClaw setup guide →

ama_retrievesemantic recall

ama_process_turnmemory lifecycle

ama_process_assistantresponse capture

ama_memory_listmemory statistics

ama_session_endepisode synthesis

ama_memory_forgetuser-controlled deletion

07 / Citation

Build on AMA.

Weiquan Huang, Zixuan Wang, Hehai Lin, Sudong Wang, Bo Xu, Qian Li, Beier Zhu, Linyi Yang, and Chengwei Qin. 2026.

AMA: Adaptive Memory via Multi-Agent Collaboration.
Findings of the Association for Computational Linguistics: ACL 2026, pages 3099–3120.

DOI ↗ Hugging Face ↗ BibTeX ↗

Memory that adaptsas dialogue evolves.