MemSkill GitHub
NeurIPS 2026Agent memory, reimagined

MemSkillLearning and Evolving Memory Skills for Self-Evolving Agents

Memory is a skill. Let agents learn it.
A framework that learns which memory skills to use, and evolves the skills themselves through experience.

Read the paper Explore code Cite this work
A self-evolving memory systemFig. 00
The MemSkill learning and evolution loop The controller selects from a shared skill bank. The executor applies those skills to construct memories. Task feedback trains the controller; hard cases guide the designer to refine and expand the skill bank. SELECTED SKILLS HARD CASES REFINE + EXPAND SHARED SKILL BANK INSERT UPDATE + Evolved skills 01 Controller Learn which skills to use 02 Executor Construct skill-guided memories 03 Designer Evolve skills from hard cases

The controller learns to select relevant skills for each text span and its retrieved memories.

Learnable skills. Reusable knowledge. Evolving memory.Explore the research
Haozhen Zhang1Quanyu Long1Jianzhu Bao1Tao Feng2Weizhi Zhang3Haodong Yue4Wenya Wang1
NTU crest

1Nanyang Technological
University

Illinois logo

2University of Illinois
Urbana-Champaign

UIC logo

3University of Illinois
Chicago

Tsinghua crest

4Tsinghua
University

01 /The idea

Better memory starts
with better ways
to remember.

Read the full abstract

Most agent memory systems rely on fixed, hand-designed operations. MemSkill introduces a different approach: learnable, reusable memory skills that guide how an agent extracts, consolidates, and revises information.

Instead of applying a fixed procedure after every turn, MemSkill processes larger spans of interaction history and composes a small set of relevant skills to construct memory in one pass.

The skill bank is a living component. Skills are refined and expanded from hard cases, allowing the system to improve both how it uses skills and what those skills are. The result is adaptable memory across long conversations, document-based reasoning, and embodied interaction.

01

Learn what to use

Select context-relevant skills from a shared bank, guided by downstream task feedback.

02

Remember in one pass

Compose selected skills over a text span to produce structured memory updates.

03

Get better with experience

Turn recurring failures into refined skills and new ways of constructing memory.

FIG. 01 / A NEW MEMORY PARADIGMExpand figure
Prior approaches repeatedly interleave turn-level handcrafted operations and LLM calls. MemSkill selects reusable skills and constructs memories from a larger text span in one pass.
From fixed operations to reusable skills. A shift from turn-level, handcrafted processing to span-level, skill-conditioned memory generation.
02 /The framework

Three components.
One evolving system.

A controller, an executor, and a designer work together to improve memory construction in a closed loop.

01 / SELECT

Controller

Uses the current text span and retrieved memories to select a small set of relevant skills. Task rewards improve its selection policy.

Reinforcement learning
02 / CONSTRUCT

Executor

Applies the selected skills with an LLM to produce structured memory updates in a single step, span by span.

Skill-conditioned generation
03 / EVOLVE

Designer

Reviews representative hard cases, refines existing skills, and proposes new ones to expand the shared skill bank.

Evolution through feedback
FIG. 02 / THE MEMSKILL ARCHITECTUREExpand figure
MemSkill architecture: a controller selects Top-K skills for an executor to update a trace-specific memory bank. Task rewards optimize the controller, while a designer mines a hard-case buffer to evolve the shared skill bank.
A closed loop of learning and evolution. The memory bank stores trace-specific information; the shared skill bank stores reusable guidance for how to remember.
03 /The evidence

Stronger memory.
Across settings.

From conversational recall to embodied tasks. Trained with LLaMA, with skills transferred directly to Qwen without additional training.

LoCoMoLLM JUDGE
53.82

+8.27% relative improvement vs. best compared baseline

LongMemEvalLLM JUDGE
60.89

+7.88% relative improvement vs. best compared baseline

ALFWorldAVG. SUCCESS RATE
80.36 percent

+3.26% relative improvement vs. best compared baseline

LLaMA 3.3 70B-Instruct · Main comparison
MethodLoCoMoLongMemEval †ALFWorld ‡
F1 ↑L-J ↑F1 ↑L-J ↑SEEN SR ↑UNSEEN SR ↑
No-Memory————62.1473.88
CoN30.8641.7230.7856.4475.0080.60
ReadAgent28.6338.2524.4842.6262.8671.64
MemoryBank36.8044.4330.5641.9660.7166.42
A-MEM39.3949.7125.8338.0462.8670.15
Mem025.4834.5830.2546.8174.2981.34
LangMem30.9135.8218.3624.3572.8679.85
MemoryOS41.3948.6417.5939.8357.8665.67
MemSkillOURS44.2153.8231.1260.8977.1483.58

Metrics. F1: token-level F1; L-J: LLM-judge score; SR: success rate (%). ↑ Higher is better; — not reported. Bold marks the best score per column.

Evaluation. † MemSkill on LongMemEval uses LoCoMo-trained skills without further training. Qwen results transfer from LLaMA without retraining. ‡ ALFWorld uses in-context demonstrations.

Source: paper v2, Table 1 ↗

Beyond the training distribution

Skills that travel
beyond conversation.

The LoCoMo-trained skill bank transfers directly to HotpotQA: a shift from dialogue to document-centric QA. Compare 50, 100, and 200 concatenated documents, and explore how the number of selected skills affects performance.

See the transfer study
HotpotQA · LLM-judge score
MemSkillMemoryOSA-MEM
MemSkill: 67.57
MemoryOS: 66.02
A-MEM: 65.63

50 documents

MemSkill: 71.48
MemoryOS: 67.19
A-MEM: 66.80

100 documents

MemSkill: 67.97
MemoryOS: 60.55
A-MEM: 61.72

200 documents

Y-axis: 55–75 · LLaMA 3.3 70B · No HotpotQA training · Paper v2, Fig. 3

04 /Inside the skill bank

Learning how
to remember.

Evolved skills capture reusable instructions for constructing memory. Explore behaviors that emerge from conversation and embodied tasks.

SKILL 01 / TEMPORAL CONTEXT

Capture temporal context

Purpose
Preserve the dates, times, durations, and sequence of events or facts.
When to use
A text span mentions an event or activity with associated temporal information.
How to apply
Identify the key temporal elements and record them concisely, keeping relevant ordering relationships.
Constraint
Use explicitly stated time information. Avoid inferring dates or durations that the text does not provide.
SKILL 02 / ACTIVITY DETAILS

Capture activity details

Purpose
Keep the context that makes an activity specific and useful to retrieve later.
When to use
An activity or event appears with details about its participants, location, or context.
How to apply
Extract the activity type, location, participants, timing, and relevant contextual information.
Constraint
Keep details specific and concise. Preserve what is stated without inventing missing context.

Adapted from the evolved skills in the paper. Each skill specifies when and how to remember.

View the original case study
05 /Explore & build

Build on MemSkill.

Read the research, explore the implementation, and experiment with learned memory skills.

The paper

Method, experiments & analysis

The code

Implementation & getting started

The checkpoints

Pretrained controller weights

Cite this work

If MemSkill is useful in your research, please consider citing our paper.

BibTeX / memskill.bib
@article{zhang2026memskill,
  title   = {MemSkill: Learning and Evolving Memory Skills
             for Self-Evolving Agents},
  author  = {Zhang, Haozhen and Long, Quanyu and Bao, Jianzhu
             and Feng, Tao and Zhang, Weizhi and Yue, Haodong
             and Wang, Wenya},
  journal = {arXiv preprint arXiv:2602.02474},
  year    = {2026}
}

Research figure

Open original