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.

Weiquan Huang^1,*, Zixuan Wang^1,*, Hehai Lin¹, Sudong Wang¹, Bo Xu¹, Qian Li², Beier Zhu³, Linyi Yang⁴, Chengwei Qin^1,†

¹HKUST (Guangzhou) · ²Shandong University · ³Nanyang Technological University · ⁴SUSTech

^*Equal contribution · ^†Corresponding author

Read the paper ↗ Explore the code ↗ arXiv:2601.20352

0.774LoCoMo LLM score

0.698LongMemEval average

≈80%fewer input tokens*

Research reference code

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

View research code ↗

Harness Agent memory skill

A portable AMA memory skill for agent-harness architectures, plus an OpenClaw adapter as a runnable reference integration.

See the agent skill ↓

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 · all reported baselines

Method	GPT-4o-mini	GPT-4.1-mini	Qwen3-30B	Qwen3-8B
FullContext	0.717	0.786	0.733	0.696
RAG	0.300	0.321	0.307	0.298
LangMem	0.515	0.717	0.525	0.512
MemGPT	0.415	0.425	—	—
Zep	0.580	0.601	—	—
A-Mem	0.476	0.487	0.476	0.464
Mem0	0.608	0.647	—	—
Nemori	0.740	0.774	0.756	0.686
AMA	0.774	0.805	0.791	0.707

Dashes indicate evaluations omitted for commercial systems under open-source backbones, following the paper's experimental protocol.

LongMemEval_sAccuracy · all reported baselines

Question type	Full-ctx	RAG	LangMem	A-Mem	MemGPT	Mem0	Zep	Nemori	AMA
Single-session preference	0.300	0.333	0.267	0.367	0.200	0.333	0.533	0.467	0.467
Single-session assistant	0.818	0.714	0.777	0.804	0.857	0.875	0.750	0.839	0.964
Temporal reasoning	0.365	0.280	0.353	0.398	0.451	0.399	0.541	0.617	0.444
Multi-session	0.406	0.254	0.424	0.451	0.549	0.481	0.474	0.511	0.624
Knowledge update	0.769	0.385	0.578	0.602	0.410	0.654	0.744	0.615	0.897
Single-session user	0.814	0.686	0.750	0.750	0.814	0.857	0.929	0.886	0.986
Average	0.548	0.374	0.528	0.538	0.554	0.574	0.632	0.642	0.698

Cross-backbone consistency. AMA achieves the highest overall LoCoMo LLM Score with all four tested backbones, from Qwen3-8B-Instruct to GPT-4.1-mini.

Long-term updates. On LongMemEval_s, AMA leads the average and reaches 0.897 on knowledge-update questions and 0.986 on user-specific recall.

Transparent limitations. Nemori remains stronger on temporal reasoning and Zep on single-session preferences, leaving clear room 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.

For Harness Agents

AMA, packaged as a portable memory skill.

The repository translates AMA into a reusable SKILL.md for agent-harness architectures. It specifies when an agent should retrieve, capture, refresh, summarize, inspect, or forget memory while preserving the framework's three complementary granularities.

Runtime integration: the skill is harness-oriented and portable; the bundled OpenClaw plugin is a runnable reference adapter that contributes automatic recall/capture hooks, six explicit tools, and a local SQLite/FAISS sidecar.

reference adapter · OpenClaw

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

Read the memory skill ↗ OpenClaw adapter 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.

BibTeX

@inproceedings{huang-etal-2026-ama,
  title     = {{AMA}: Adaptive Memory via Multi-Agent Collaboration},
  author    = {Huang, Weiquan and Wang, Zixuan and Lin, Hehai and
               Wang, Sudong and Xu, Bo and Li, Qian and Zhu, Beier and
               Yang, Linyi and Qin, Chengwei},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026},
  pages     = {3099--3120},
  doi       = {10.18653/v1/2026.findings-acl.152}
}

DOI ↗ Hugging Face ↗ Citation file ↗

Memory that adaptsas dialogue evolves.