Source-linked AI summary
AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution
Yutao Yang, Junsong Li, Qianjun Pan, Bihao Zhan, Yuxuan Cai, Lin Du, Jie Zhou, Kai Chen, Qin Chen, Xin Li, Bo Zhang, Liang He
TL;DR
LLM agents often fail to consolidate users’ recurring preferences and requirements into reusable capabilities across sessions. AutoSkill converts interaction experience into explicit, editable skills that are retrieved and refined at inference time without retraining, providing a framework for persistent personalized behavior.
Problem
Recurring user preferences, constraints, and workflows are often treated as retrievable text rather than operationalized behavior, leaving a mechanism for explicit reusable skills missing.
Method
AutoSkill extracts, represents, versions, retrieves, merges, and refines explicit skill artifacts through a model-agnostic plugin layer without optimizing model parameters.
Results
AutoSkill provides a practical framework that accumulates recurring preferences, stylistic requirements, and workflows as explicit, reusable, maintainable behavioral knowledge without retraining the underlying model.
Takeaways & Limitations
Editable skill artifacts support transparency, controllability, deployability, and compatibility with existing models and agent systems.
Takeaways & Limitations
The demonstrated skill artifacts encode task-specific operational constraints, such as explicit waiting, placeholder, and task-separation rules, rather than a single universal procedure.
Abstract
from arXiv · showhide
In practical LLM applications, users repeatedly express stable preferences and requirements, such as reducing hallucinations, following institutional writing conventions, or avoiding overly technical wording, yet such interaction experience is seldom consolidated into reusable knowledge. Consequently, LLM agents often fail to accumulate personalized capabilities across sessions. We present AutoSkill, an experience-driven lifelong learning framework that enables LLM agents to automatically derive, maintain, and reuse skills from dialogue and interaction traces. AutoSkill abstracts skills from user experience, supports their continual self-evolution, and dynamically injects relevant skills into future requests without retraining the underlying model. Designed as a model-agnostic plugin layer, it is compatible with existing LLMs and introduces a standardized skill representation for sharing and transfer across agents, users, and tasks. In this way, AutoSkill turns ephemeral interaction experience into explicit, reusable, and composable capabilities. This paper describes the motivation, architecture, skill lifecycle, and implementation of AutoSkill, and positions it with respect to prior work on memory, retrieval, personalization, and agentic systems. AutoSkill highlights a practical and scalable path toward lifelong personalized agents and personal digital surrogates.
1 Introduction
AutoSkill addresses the missing mechanism for converting recurring interaction experience into explicit, reusable skills. It represents, maintains, and reuses those skills as editable artifacts without retraining the underlying model.
- Recurring user preferences and operating requirements are often repeated across sessions but not consolidated into reusable knowledge.
- AutoSkill abstracts behaviors such as stylistic constraints, response strategies, tool procedures, and domain conventions into structured skill artifacts.
- Editable and versioned skills support inspection, revision, merging, transparency, controllability, and long-term maintenance.
- Its lifecycle extracts candidate skills, summarizes them into SKILL.md artifacts, refines them, and injects relevant skills into future requests.
- AutoSkill is presented as an open-source, deployable plugin layer compatible with existing language models and agent pipelines.
2 Related Work
Related work spans experience-driven lifelong learning, self-evolution, long-term memory, and skill learning for reasoning and acting. AutoSkill distinguishes itself through explicit, editable skill artifacts and lifecycle management.
- AutoSkill emphasizes explicit skill artifacts, human-editable representations, and lifecycle management for extraction, revision, retrieval, and reuse.
- Experience-driven lifelong learning studies how agents accumulate reusable knowledge, strategies, or policies from ongoing interactions.
- Self-evolution methods improve behavior through self-reflection, iterative rewriting, feedback refinement, or autonomous data construction.
- Memory-based approaches retrieve external knowledge or stored interaction information to support factuality, traceability, and future generation.
- Skill learning for LLM agents targets reusable reasoning patterns, tool-use procedures, and action strategies.
3 Method
AutoSkill uses two coupled loops: one retrieves skills to enhance responses, while the other extracts and maintains skills from interaction experience. The framework improves through explicit memory growth rather than fine-tuning.
- The response-generation loop retrieves useful skills, while the evolution loop updates the skill bank from newly observed dialogue turns.
- Skill-enhanced response generation uses query rewriting, skill retrieval, and context injection to support the current response.
- Skill evolution transforms interaction experience into explicit skills through extraction and maintenance.
- The coupled design enables continual improvement through explicit memory growth without modifying model parameters.
3.1 Problem Definition
For each user, AutoSkill maintains dialogue history and a user-specific skill bank, then uses prompt-driven modules to retrieve skills and update that bank. Skill extraction relies on user queries rather than model responses.
- The complete dialogue history for user u is represented as X_u = {x_1, x_2, . . . , x_T}, with each turn containing a query and response.
- A user-specific skill bank B_t^u is maintained after turn t.
- Each skill includes a name, description, executable prompt, trigger set, tag set, example set, and version.
- The training-free method uses prompt-driven modules for query rewriting, dialogue generation, skill extraction, management decisions, merging, and embedding-based vectorization.
- Skill extraction learns from user queries {q_1, . . . , q_t} and does not use model responses as extraction evidence.
3.2 Prompt-Driven Modular Architecture
AutoSkill implements its functions as a modular inference-time composition of task-specific prompts paired with general-purpose models. This modularity allows components and backbone models to be replaced without retraining the framework.
- Modular composition: Task-specific prompts instantiate query rewriting, dialogue generation, skill extraction, management decisions, and skill merging as modular functions.Each module is paired with a general-purpose LLM; the embedding model supports dense vector retrieval.
- Modular composition: Different modules can share one backbone LLM while serving distinct roles through different prompts.
- Model flexibility: Replacing the response, extraction, or embedding model does not require retraining the framework.
3.3 Skill-Enhanced Response Generation
The response loop rewrites the current query, retrieves relevant skills through hybrid semantic and lexical matching, and injects threshold-qualified skills into generation. If no skill qualifies, the model answers without augmentation while keeping its parameters unchanged.
- Query rewriting: Query rewriting produces a concise retrieval-oriented query that preserves task anchors and exposes format, style, structure, or domain constraints.
- Hybrid retrieval: AutoSkill combines normalized dense semantic and lexical BM25 scores using λ to balance semantic and exact matching.
- Skill selection: Skills are ranked by relevance, and only the top-K candidates above threshold η are injected into the dialogue context.
- Skill-conditioned generation: The response generator uses retrieved skills only when they directly match the user’s intent and otherwise answers normally.
- Skill-conditioned generation: The response model adapts to accumulated user-specific experience while its model parameters remain unchanged.
3.4 Real-Time Skill Evolution
AutoSkill extracts durable, portable requirements from recent user queries, compares candidates with nearby existing skills, and adds, merges, or discards them. Merging preserves skill identity while versioned refinement incorporates reusable new constraints without duplicating skills.
- Skill extraction: Skill extraction uses recent user queries to identify durable preferences, procedures, constraints, policies, and templates rather than one-off requests.
- Skill extraction: The extractor abstracts portable rules from recurring interactions while removing case-specific entities and assistant-invented details.
- Skill representation: Extracted skills use structured fields including a name, description, prompt, triggers, tags, examples, and confidence.
- Skill management: A management judge compares each candidate with its most similar existing skill and selects add, merge, or discard.
- Skill management: Nearest-neighbor management makes decisions more focused and scalable than reasoning over the entire skill bank.
- Versioned skill merging: Merging preserves the existing capability identity while integrating reusable, non-conflicting constraints, examples, or execution details into an updated version.
- Versioned skill merging: Version iteration lets the same skill track evolving requirements and prevents duplicated skills when feedback arrives on a recurring task.
3.5 Training-Free Lifelong Learning
AutoSkill realizes lifelong learning through an external skill-memory loop combining response-time retrieval with after-turn skill evolution. The framework is training-free and prompt-driven because it constructs, retrieves, and refines explicit skills without optimizing model parameters.
- External skill memory: The response loop combines query rewriting, hybrid retrieval, thresholded Top-K injection, and skill-conditioned generation.
- External skill memory: The evolution loop performs query-only extraction, nearest-neighbor management, and versioned merging after each turn.
- Training-free learning: No model parameters are optimized; improvements come from explicit skill construction, retrieval, and refinement.
4 System Overview
AutoSkill is a lifelong learning layer that transforms recurring interactions into explicit, maintainable skills and reuses them during future requests. Its architecture separates foreground retrieval and generation from background skill extraction and maintenance, enabling controlled, inspectable capability accumulation.
- System Overview: AutoSkill transforms recurring preferences, constraints, and workflows into persistent skill artifacts retrieved and injected into future requests.The design separates an online serving path from a background learning path.
- Implementation and Usage: AutoSkill uses persistent SkillBank storage, vector indexing, and SDK, Web UI, and OpenAI-compatible proxy interfaces for deployment across development and service settings.The storage organization separates user-specific and shared skills, while the interfaces expose ingestion, search, rendering, retrieval, and context injection.
- Skill Abstraction Layer: Skills are represented as SKILL.md-centered artifacts containing identity, metadata, executable instructions, and optional supporting resources.This makes learned behavior a first-class object that can be reviewed and maintained explicitly.
- Skill Management Layer: The management layer extracts candidates from interaction traces, then adds, merges, or discards them based on comparison with the existing repository.Discarding noisy or one-off patterns supports controlled repository growth.
- Skill Lifecycle: The four-stage lifecycle ingests experience, extracts reusable capabilities, maintains versioned skills, and retrieves them for reuse in later tasks.Later feedback updates existing artifacts rather than creating overlapping prompt fragments, preserving consistency over time.
- Serving and Interaction: At inference time, AutoSkill refines queries, retrieves and filters relevant skills, injects them into generation context, and evolves skills concurrently in the background.This keeps retrieval and response generation on the latency-critical path while extraction and maintenance proceed asynchronously.
5 Experimental Analysis
The study analyzes four multilingual SkillBank subsets and finds broad coverage of technical, writing, communication, and platform-specific capabilities. Case studies show that extracted skills are explicit artifacts that can be versioned and refined over repeated interactions.
- Dataset and SkillBank Statistics: The four subsets cover Chinese and English interactions from GPT-3.5 and GPT-4, with English GPT-3.5 largest and Chinese GPT-4 smallest.English GPT-3.5 contains 10,243 conversations, 267,681 messages, and 631 skills; Chinese GPT-4 contains 1,145 conversations, 36,834 messages, and 224 skills.
- SkillBank Statistics: Extracted skills are concentrated in programming and software development while also covering writing, communication, data, education, translation, and roleplay.Python is the most frequent normalized tag, followed by JavaScript, Excel, C++, and pandas.
- SkillBank Statistics: Skill versions provide evidence of iterative refinement at different rates across skills.The English professional_text_rewrite skill reaches version 0.1.34 after 34 optimization rounds, whereas the Chinese 顶级心理咨询师 skill remains at version 0.1.0.
- Case Studies: Two case studies demonstrate that the same structured format can encode multilingual, user-aligned capabilities across different domains and interaction styles.The examples include a Chinese counseling-oriented behavior and an English professional rewriting capability.
- Case Studies: The counseling skill stores warmth, empathy, professionalism, privacy respect, and restrictions against inappropriate medical diagnosis or drug recommendations.It converts a recurring conversational preference into an explicit artifact that can be reactivated in future psychologically supportive dialogue.
- Case Studies: The professional rewriting skill preserves meaning and factual details while prohibiting explanations, omissions, additional commentary, and multiple rewrite options.Its version 0.1.34 reflects repeated consolidation of feedback into one reusable artifact rather than duplicated prompt fragments.
6 Conclusions and Future Work
AutoSkill transforms recurring interaction experience into explicit, reusable, and maintainable skills without retraining the underlying model. The framework combines extraction, representation, retrieval, reuse, and iterative refinement to support personalized agents across languages, model families, and task domains.
- Conclusions and Future Work: AutoSkill enables lifelong learning by externalizing recurring interaction experience as explicit skill artifacts without modifying base model parameters.Its lifecycle includes skill extraction, representation, retrieval, reuse, and iterative refinement.
- Conclusions and Future Work: The framework preserves user preferences, stylistic requirements, and recurring workflows as operational behavioral knowledge.The skills remain explicit and editable, supporting transparency, controllability, and deployability.
- Conclusions and Future Work: The analysis and experiments indicate that AutoSkill can accumulate diverse capabilities from real-world interactions across languages, model families, and task domains.The paper positions external skill evolution as a scalable path toward lifelong personalized agents.
技能卡片：八字命理大师角色扮演
This skill card specifies a Chinese fortune-telling role with structured inputs, broad traditional-astrology outputs, stylistic requirements, and an entertainment disclaimer. The surrounding SkillBank examples also illustrate reusable constraints for social-media writing, roleplay, and Selenium automation.
- 八字命理大师角色扮演: The fortune-telling card requires birth date, time, gender, and location, then specifies analysis of命宫,五行,十神,纳音,神煞, and related fortunes.It also requests explanations of technical terms, practical suggestions, and a clear entertainment-only disclaimer.
- 八字命理大师角色扮演: The card explicitly prohibits claiming absolute accuracy and states that analyses are for entertainment rather than real-world guidance.This boundary is repeated as an operational requirement and notice.
- 八字命理大师角色扮演: Its communication style is professional, mysterious, directive, and approachable while explaining specialized terminology in accessible language.The card frames the behavior within Chinese traditional culture and命理学 principles.
- Related Skill Cards: A separate Xiaohongshu skill targets young women with cute, lively, emoji-heavy posts and requires explicit user instruction before generation.Its workflow separates style learning from content generation and requires title, body, and hashtag sections.
- Related Skill Cards: A cat-roleplay skill encodes mandatory “喵” and “主人” prefixes or suffixes together with persistent character consistency.These constraints define both response formatting and persona maintenance.
- Related Skill Cards: The Selenium automation skill generates Python workflows for navigation, JavaScript execution, waits, dynamic-content stabilization, and cleanup.It forbids fixed sleep-based waiting, hardcoded inputs, and unrelated tasks in one script.