Memory Papers
Memory Architecture

MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation

AuthorsHuawei Lin, Peng Li, Jie Song et al.

arXiv 20262026
arXivPDF

TL;DR

MUSE-Autoskill uses a full skill lifecycle with skill-level memory and unit-test-driven refinement to reach 59.67% on SkillsBench and 72.0% on SkillLearnBench, the best among GPT-5.5-backed agents.

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

Skills as static artifacts limit long term improvement (85.24% vs 81.17% on covered tasks)

Existing skill systems treat skills as isolated, static artifacts, which limits reusability, reliability, and long term improvement across tasks.

Without lifecycle management and per skill memory, agents suffer from creation usage mismatch, lack of validation, and poor context handling on complex, multi step workflows.

HOW IT WORKS

MUSE-Autoskill Agent Lifecycle

MUSE-Autoskill centers a Master Agent, Skill Creator, Skill Bank, Evaluator, and multi level Memory around a unified skill lifecycle of creation, memory, management, evaluation, and refinement.

You can think of MUSE-Autoskill like an operating system with RAM, disk, and a package manager, where skills are installable, testable apps with their own logs.

This lifecycle lets MUSE-Autoskill accumulate per skill experience, run unit tests, and refine skills over time, instead of relying on a plain, truncated context window.

DIAGRAM

Skill Lifecycle Flow in MUSE-Autoskill

This diagram shows how MUSE-Autoskill creates, evaluates, stores, and refines skills during task solving.

DIAGRAM

SkillsBench and SkillLearnBench Evaluation Pipeline

This diagram shows how MUSE-Autoskill and baselines are evaluated on SkillsBench and SkillLearnBench.

PROCESS

How MUSE-Autoskill Handles a Task — Skill Creation, Memory, Management, Evaluation, Refinement

  1. 01

    Planning

    MUSE-Autoskill uses the Master Agent to interpret the query, decompose the problem, and decide whether to invoke skills from the Skill Bank or call the Skill Creator.

  2. 02

    Action

    MUSE-Autoskill invokes existing skills or built in tools like skill_create and web_search, executing scripts via the Skill Executor inside sandbox containers.

  3. 03

    Observation

    MUSE-Autoskill collects execution results and artifacts, feeding them back into the ReAct loop and appending notes into short term, long term, and skill level Memory.

  4. 04

    Skill Evaluation and Refinement

    MUSE-Autoskill runs unit tests with the Evaluator, triggers the Refiner on failures, and updates or merges skills in the Skill Bank while persisting per skill experience.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Unified Skill Lifecycle

    MUSE-Autoskill formalizes a five stage lifecycle—creation, memory, management, evaluation, refinement—covering all stages where prior systems like Voyager and AutoSkill only partially operate.

  • 02

    Skill Level Memory and Context Management

    MUSE-Autoskill adds skill level .memory.md files plus short term and long term Memory, and uses a DAG based context manager with two level compression to handle long horizon tasks.

  • 03

    Best in Class SkillsBench and SkillLearnBench Results

    MUSE-Autoskill reaches 59.67% accuracy with human skills and 53.42% with self created skills on SkillsBench, and 72.0% with human skills and 48.0% self created on SkillLearnBench.

RESULTS

By the Numbers

SkillsBench Accuracy

59.67%

+12.72 pp over MUSE-Autoskill without skills (46.95%)

SkillLearnBench Accuracy

72.0%

+29.0 pp over MUSE-Autoskill without skills (43.0%)

Self Created SkillsBench

53.42%

+6.47 pp over MUSE-Autoskill without skills (46.95%)

Hermes Transfer Accuracy

51.90%

+14.66 pp over Hermes without skills (37.24%) and +3.88 pp over Hermes with human skills (48.02%)

SkillsBench is a 75 task benchmark of real world, Docker graded tasks, while SkillLearnBench has 20 tasks and 100 verified instances for continual skill generation. These results show that MUSE-Autoskill’s lifecycle managed skills improve both its own performance and cross agent transfer beyond human authored skills.

BENCHMARK

By the Numbers

SkillsBench is a 75 task benchmark of real world, Docker graded tasks, while SkillLearnBench has 20 tasks and 100 verified instances for continual skill generation. These results show that MUSE-Autoskill’s lifecycle managed skills improve both its own performance and cross agent transfer beyond human authored skills.

BENCHMARK

SkillsBench Main Accuracy Results on 75 Task Common Set

Accuracy (%) on SkillsBench with human skills for GPT-5.5 backed agents.

BENCHMARK

SkillLearnBench Accuracy Results on 100 Verified Instances

Accuracy (%) on SkillLearnBench with human skills for GPT-5.5 backed agents.

KEY INSIGHT

The Counterintuitive Finding

On covered SkillsBench tasks, MUSE-Autoskill’s self created skills reach 85.24% accuracy, beating human authored skills at 81.17% on the same subset.

This is surprising because we usually expect human skills to be stronger, yet lifecycle managed, experience aware skills distilled from trajectories become more effective and cheaper to reuse.

WHY IT MATTERS

What this unlocks for the field

MUSE-Autoskill shows that agents can treat skills as long lived, testable infrastructure with skill level memory, rather than disposable prompts.

Builders can now create agents that continuously accumulate and refine reusable capabilities, share skills across systems like Hermes, and amortize exploration cost over many future tasks.

~14 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer

Questions about this paper?

Paper: MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation

Answers use this explainer on Memory Papers.

Checking…