AuthorsYutao Yang, Junsong Li, Qianjun Pan et al.
TL;DR
AutoSkill turns recurring interaction experience into explicit versioned skills via query-only extraction and hybrid retrieval, enabling training-free lifelong personalization with a growing SkillBank of 1,858 skills from WildChat-1M.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
AutoSkill targets the failure that interaction experience is seldom consolidated into reusable knowledge, so user habits must be reestablished from scratch in each session.
This breaks long term personalization for writing assistance, planning, and coding, because agents never crystallize stable preferences, constraints, or workflows into durable capabilities.
HOW IT WORKS
AutoSkill’s core mechanism chains Query Rewriting, Hybrid Skill Retrieval, Skill Extraction, Skill Management Decision, and Versioned Skill Merging around a persistent SkillBank of SKILL.md artifacts.
You can think of AutoSkill like a personal library card catalog layered beside an LLM: each recurring behavior becomes a labeled, versioned card instead of a forgotten conversation.
This explicit skill memory lets AutoSkill preserve stable styles, policies, and workflows across sessions, something a plain context window or raw retrieval log cannot reliably maintain.
DIAGRAM
This diagram shows how AutoSkill rewrites queries, retrieves skills with hybrid dense plus BM25 scoring, and injects them into skill-conditioned response generation.
DIAGRAM
This diagram shows how AutoSkill ingests experience, extracts skill candidates, applies add or merge decisions, and updates the SkillBank with versioned skills.
PROCESS
Experience ingestion
AutoSkill ingests dialogue history X_u and user query sequence Q_tu so Skill Extraction can see recent user-only evidence for durable preferences and workflows.
Skill Extraction from Interaction
Using Skill Extraction, AutoSkill proposes candidates z_t with fields like name, description, prompt, triggers, tags, examples, and confidence from recent user queries.
Retrieval-Assisted Skill Management
AutoSkill applies Skill Management Decision by retrieving nearest neighbors from the SkillBank and choosing add, merge, or discard for each candidate.
Versioned Skill Merging
Through Versioned Skill Merging, AutoSkill semantically merges candidates into existing skills, bumps version numbers, and updates the SkillBank without duplicating artifacts.
KEY CONTRIBUTIONS
Experience-driven lifelong learning framework
AutoSkill formulates transforming interaction experience into explicit reusable skills and operationalizes this with Skill Extraction and Skill Management Decision over user-specific SkillBanks built from WildChat-1M.
Skill lifecycle with self-evolution
AutoSkill introduces a full lifecycle covering Query Rewriting, Hybrid Skill Retrieval, Skill Extraction, and Versioned Skill Merging, enabling continual evolution without any parameter updates.
Editable SKILL.md artifacts and deployable system
AutoSkill designs skills as editable SKILL.md artifacts with version fields like 0.1.34 and ships an open source system with SDK, Web UI, and OpenAI-compatible proxy for practical deployment.
RESULTS
Conversations count
10243 conversations
+4331 over Chinese GPT-3.5 subset
Total messages
267681 messages
vs English GPT-4 subset
Extracted skills
631 skills
SkillBank size for English GPT-3.5 subset
Avg messages per conversation
32.17 messages
highest average in Chinese GPT-4 subset
AutoSkill is evaluated by building four SkillBank subsets from WildChat-1M, measuring conversations, messages, and extracted skills to show that AutoSkill can induce 1,858 explicit skills from long real-world dialogues.
BENCHMARK
AutoSkill is evaluated by building four SkillBank subsets from WildChat-1M, measuring conversations, messages, and extracted skills to show that AutoSkill can induce 1,858 explicit skills from long real-world dialogues.
BENCHMARK
Extracted Skills per subset in AutoSkill’s SkillBank.
KEY INSIGHT
AutoSkill shows that a single English skill, professional_text_rewrite, can reach version 0.1.34 through incremental merges without any model fine-tuning.
This is surprising because many assume stable personalization requires parameter updates, yet AutoSkill achieves rich evolution entirely via external SKILL.md artifacts.
WHY IT MATTERS
AutoSkill unlocks training-free lifelong personalization where agents accumulate explicit, inspectable skills instead of opaque latent adaptations.
Builders can now ship agents that remember styles, workflows, and policies across sessions via SKILL.md artifacts, while keeping base models frozen and easily swappable.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…