⌘K
Change language Switch ThemeSign In
Narrow Mode
MetaClaw: An In-depth Analysis of a Self-Evolving AI Agent Framework Based on Real Conversations ================================================================================================
MetaClaw: An In-depth Analysis of a Self-Evolving AI Agent Framework Based on Real Conversations ================================================================================================  ### meng shao
@shao__meng
MetaClaw: "在真实对话中自我进化"的 AI Agent 框架
项目地址：github.com/aiming-lab/Met…
项目定位
在用户与 LLM 的日常对话之间插入透明代理，将每次交互转化为学习信号，通过 Skills 注入和在线 RL 两条通路持续提升 Agent，无需自建 GPU。
架构四层
· 代理层 — FastAPI 伪装为 OpenAI 兼容 API，拦截 OpenClaw 请求，注入 Skills 后转发上游 LLM，同时记录对话供训练使用。
· Skills 系统 — SkillManager 管理 Markdown Skills 库（九大类别，支持关键词匹配和语义嵌入两种检索）；SkillEvolver 分析失败样本后调用 LLM 自动生成新 Skills 写回库中。Skills 库维护 generation 计数器，新 Skills 写入即递增。
· RL 训练器 — GRPO 算法在线微调：代理采样 → PRM 异步评分 → Tinker/MinT 云端 LoRA 训练 → 热替换权重。可选 OPD 蒸馏。推理与训练全程异步不阻塞。
· 元学习调度器 — 将 RL 权重更新推迟到睡眠时段、键盘空闲或日历会议期间执行。四态状态机（IDLE_WAIT → WINDOW_OPEN → UPDATING → PAUSING），用户中途回来则暂存 batch 下次继续。
关键设计：MAML 式 Support/Query 分离
学习分为快速内循环（技能进化，会话后立即执行）和慢速外循环（RL 梯度更新，空闲时执行）。技能进化导致 generation 递增时，训练器清空所有旧样本——触发进化的数据（support set）不会被重复用于 RL 更新（query set）。这是 MAML 思想在在线部署场景的工程落地。
三种模式
· skills_only：LLM API，技能注入 + 自动总结，无需 GPU
· rl：+ Tinker/MinT，持续 RL，batch 满即训
· madmax（默认）：+ 调度器，RL + 空闲调度 Show More
#### Cihang Xie
@cihangxie · 5d ago
Super excited to present MetaClaw 🦞 — our self-evolving agent that learns just by talking to it.
💬 Conversations → training trajectories
🧠 Failures → new reusable skills
⚡ Weights update live (no interruption)
☁️ Powered 100% via API — no GPU cluster needed
Try it heregithub.com/aiming-lab/Met…mwzR
Huge kudos to the team @richardxp888, Jianwen Chen, @Xinyu2ML, @lillianwei423, @StephenQS0710, Zeyu Zheng, @HuaxiuYaoMLShow More
00:18
5
25
92
10.1K
Mar 14, 2026, 1:14 PM View on X
0 Replies
1 Retweets
3 Likes
324 Views  meng shao @shao__meng
One Sentence Summary
MetaClaw is an Agent framework that achieves self-evolution by intercepting conversations via a transparent proxy and utilizing Skill injection and online Reinforcement Learning (RL) algorithms.
Summary
This tweet provides an in-depth analysis of MetaClaw framework's architecture and core principles. The project is positioned as "evolving through conversations," recording signals by inserting a FastAPI proxy layer between users and LLMs. Its architecture comprises a Proxy Layer, a Skills System (automatic generation and retrieval), an RL Trainer (using the GRPO algorithm for online fine-tuning), and a Meta-Learning Scheduler (updating weights during idle times). The tweet particularly highlights its MAML-style Support/Query separation design, which decouples skill evolution from gradient updates. This framework supports a continuous learning mode that is entirely API-based and requires no local GPU.
AI Score
88
Influence Score 1
Published At Today
Language
Chinese
Tags
MetaClaw
AI Agent
Reinforcement Learning
MAML
Self-Evolution HomeArticlesPodcastsVideosTweets
MetaClaw: An In-depth Analysis of a Self-Evolving AI Agen... ===============