MetaClaw Explained: Self-Improving Skills and Memory for AI Agents
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I break down MetaClaw, an open-source repo that helps your AI agent learn from the conversations you have with it. I go over how its proxy works, its three modes, its cross-session memory layer, and how to set it up with OpenClaw. It's for anyone running an AI agent for real work who's tired of it starting from scratch every session.
I recorded this in March 2026, when the repo had about 3,100 GitHub stars. Tools like this change fast, so check the current docs in the repo before you install anything.
What I cover in this video
- 0:46 The problem: your agent's underlying model doesn't learn from you
- 1:34 How MetaClaw sits as a proxy and adds skills to your prompts
- 2:19 Turning finished sessions into new skills, plus a look at RL mode
- 4:35 The three modes: skills-only, RL, and Mad Max
- 6:50 The Contexture layer for cross-session memory
- 7:35 Setting it up with OpenClaw
- 9:06 Why this matters for builders
- 10:38 The research paper behind it
The problem MetaClaw solves
If you run OpenClaw, Claude Code, or any other agent, each conversation basically starts fresh. You've got memory files, a soul markdown file, and an agents markdown file, and they help a lot. I run my own memory setup with Supermemory, daily log files, and long-term memory consolidation, and it makes a big difference compared to running an agent straight out of the box. But the model behind your agent never changes. It's the same base model everyone else is using, and all it's doing is reading your notes.
MetaClaw sits as a proxy between you and your agent. Your message goes through MetaClaw first. It looks at what you're asking, checks its skill library for anything relevant, and adds those skills and that context to your prompt before it reaches the model. You keep chatting like you always do.
When a session ends, MetaClaw summarizes what happened and turns what it learned into new skills. Say you spend 30 minutes debugging a React component and work out the steps that fixed it. MetaClaw saves that, so the next time a similar problem comes up, the skill is ready to go.
The three modes
Skills-only mode
This is the lightweight option. You don't need a GPU or a training setup, and there's no extra cost. It runs the proxy, adds skills on every turn, and turns your sessions into new skills when you're done. If you're new to AI agents, start here. It's basically plug and play.
RL mode
RL stands for reinforcement learning. Once MetaClaw has collected enough interaction data, it batches it up and trains a lightweight model on top of your interactions with LoRA fine-tuning. It doesn't train during your active sessions. It waits for idle windows, and it can connect to your Google Calendar to see when you're busy. You'll need a training backend. Right now there are three options: Tinker (the default, a cloud-based LoRA training service), MinT, and Weaver.
Mad Max mode
Yes, that's the real name. It's the full package: skills, RL, and a scheduler that uses your Google Calendar to figure out when you're in meetings, asleep, or away, then runs the weight updates during those times. Your agent stays fast while you're using it and improves in the background while you're not. This is the mode I'd eventually want to run myself, because you set it once and leave it.
The Contexture layer and OpenClaw setup
Version 0.4 added the Contexture layer, which gives MetaClaw persistent memory across sessions. It remembers facts about you, like your preferences, project history, and coding patterns, and pulls the relevant ones into your prompt on every turn. So weeks later, your agent can still use something you told it three weeks ago.
For OpenClaw, the setup is short:
- Download the plugin file from the GitHub releases page.
- Unzip it into your OpenClaw extensions folder.
- Enable it and restart your gateway.
Next, run the MetaClaw setup wizard, where you pick your agent type and model provider, and then start it. In the video I say these as "MetaClaw setup" and "MetaClaw start". Check the repo README for the exact command syntax:
metaclaw setup
metaclaw start
It doesn't only work with OpenClaw. It also supports IronClaw, PicoClaw, ZeroClaw, NanoClaw, NemoClaw, and Hermes Agent. During setup it patches the config files, restarts the right services, and connects everything for you, so you aren't locked into one ecosystem.
Why this matters, and the research behind it
Over time, a MetaClaw-enhanced agent becomes different from everyone else's. It learns your coding patterns, your debugging style, the frameworks you use, and the solutions that work in your projects. It's like the difference between hiring a new contractor every day and working with a teammate who's been with you for months. I run 13 AI agents on my Mac Mini: The Rizza writes scripts and outlines, Inspector Deck handles X, Ghostface does scraping and research, and Method Man cuts long videos into shorts. They all have memory files and personality configs, but they all run on the same underlying model as everyone else. MetaClaw could tune each one to how I actually work.
The team behind it is from the University of California, Santa Cruz. Their paper is on arXiv (2603.17187), and it hit number one on Hugging Face daily papers. According to the paper, agents using MetaClaw improve measurably over time across several benchmarks. Skill injection alone makes responses more relevant and accurate, and with RL turned on, the model weights shift toward your specific tasks. To be upfront: I hadn't deployed it myself when I made this video. I'd found the repo that day and planned to test it over the following week.
Key takeaways
- MetaClaw is a proxy that adds relevant skills and context to your prompts before they reach the model.
- It turns finished sessions into new skills automatically, so solutions you've already found get reused.
- Start with skills-only mode. It needs no GPU and costs nothing extra. Worst case, your agent gets a bit more context on each turn.
- RL and Mad Max modes add LoRA fine-tuning during idle time, but you need a training backend (Tinker, MinT, or Weaver).
- Setting it up with OpenClaw takes a few steps, and it also works with Hermes Agent and several other Claw-style agents.
FAQ
What is MetaClaw?
MetaClaw is an open-source GitHub repo from aiming-lab. It sits as a proxy between you and your AI agent, adds relevant skills to your prompts, and turns your sessions into new skills so the agent keeps improving.
Do I need a GPU to use MetaClaw?
Not for skills-only mode. It needs no GPU and has no extra cost. RL mode and Mad Max mode need a training backend such as Tinker, MinT, or Weaver.
Does MetaClaw work with OpenClaw and Hermes Agent?
Yes. For OpenClaw, you download the plugin from the releases page, unzip it into your extensions folder, enable it, restart your gateway, then run the setup wizard and start MetaClaw. It also supports Hermes Agent, IronClaw, PicoClaw, ZeroClaw, NanoClaw, and NemoClaw.
When does MetaClaw train the model?
In RL and Mad Max modes, training only runs during idle windows, never during your active sessions. It can connect to your Google Calendar to know when you're in meetings, asleep, or away.
Chapters
Transcript
Read the full transcript
0:00This GitHub repo is going to make your AI agent a thousand times smarter. It has over 3,000 stars and it is going crazy on GitHub right now. I have it linked down below. You need to check it out. So, this repo basically makes your AI agent smarter in every single way and every time you talk to it, it gets progressively better. I'm not exaggerating. This has insane value right now and there's a reason why it has over 3,100 stars on GitHub and it just dropped a research paper that hit number one on Hugging Face daily. So, people are paying attention. And when you see what it actually does, you're going to understand why it has so much value in the AI agent space. So, the repo is called Meta Claw and what it
0:46does is crazy. Right now, if you're running Open Claw or Claude code or if you have any AI agent for that matter, every conversation you have with it basically starts fresh. So, you've got memory files. You've got your soul markdown file. You have your agent's markdown file. But, the actual model, the brain behind your agent doesn't actually learn anything from talking to you. It doesn't get better. It doesn't adapt. Every single time, it's the same base model just reading your notes. Meta Claw changes that entirely. So, what it does, you know, it without you having to do anything different, you're just conversing back and forth. You know, how it works is pretty interesting. Meta Claw sits as a proxy between you and
1:34your AI agent. So, when you talk to your Open Claw agent, for example, instead of going straight into the model and talking back and forth, your message goes through Meta Claw first. And Meta Claw does something really clever here. It looks at what you're about to ask. It checks the skill library for anything relevant to your question and it injects those skills and context into your prompt before it even hits the model. So, your agent is getting smarter inputs every single time without you lifting a finger. You just chat like normal. Meta Claw handles the rest. And that's just the beginning here. The real magic happens after the conversation ends. So, when your session wraps up, Meta Claw automatically summarizes what happened and turns those insights into new
2:19skills. So, if you spend 30 minutes debugging like a React component, for example, and you figured out like the pattern and workflow that got the bug fixed, Meta Claw will capture that. So, the next time you hit a similar problem, it already knows the solution. It already has that skill ready to inject. You don't have to remember it. You don't have to write it down. Meta Claw saw you solve the problem and it learned from it. So, here's where it gets more interesting is they have a thing called RL mode. RL stands for reinforcement learning. If you enable it, Meta Claw actually trains a lightweight model on top of your interactions using a Laura fine-tuning. So, not during your active sessions, though. It's smart about it. It only runs the training during what
3:04they call idle windows. So, when you're sleeping, when you're in a meeting or when you're away from your desk, it actually connects to your Google calendar to figure out when you're busy. So, your agent is literally getting smarter while you're doing other things. So, you think about that for a second. You go to bed, you wake up, hopefully, your agent is measurably better at helping you than it was yesterday. And that's kind of what we've been looking for in this AI agent space. We've been dealing a lot with persistent memory, amnesia and dementia with our AI agents. Finally, something that fixes this problem. So, it's not science fiction anymore. Someone's actually building this problem and it's blowing up right now. And that's what the repo does right now, today. You can install it right now, today and have this into your Open
3:49Claw setup or any AI agent workflow. So, I'm going to break down the architecture because I think this is important that you understand this. Like, what's actually happening under the hood. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full Open Claw course and four live bootcamps every single week where we actually help you get set up from scratch. Like, actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you can book a call with me. We can share screens and I can help you get Claude code or Open Claw running on your machine. That's it. No fluff. I
4:35built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people who hold you accountable and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Meta Claw has three modes. So, the first one is called skills-only mode. This is the lightweight version. So, there's no GPU required, no fancy training setup, there's no extra cost. It just runs the proxy and it injects those skills at every turn and then it auto summarizes sessions into new skills when you're done. So, if you're just getting started with AI agents, this is where I'd recommend you begin first. It's basically plug and play. You
5:20install it and you start it up and then your agent your agent immediately starts building up a skill library from your conversations. The second mode is called RL mode. We talked about this earlier. It's the reinforcement learning training on top of skills. So, let's just say when Meta Claw collects enough interaction data, it batches it up and then it trains. So, you need to have a training back-end for this. They support three options right now. Tinker, which is their default cloud-based Laura training service built by Moonshot AI team. There's Mint, which is an alternative for people who want infrastructure. And then there's Weaver, which is the third option from Next AGI team. You pick up whatever which one works for your setup and you point it to
6:05whatever choice you want to do. It's up to you. The third model is called Mad Max mode and yes, that's actually what they called it. This is the third mode in in the whole Meta Claw repo. It didn't make this up. Mad Max, like what does this exactly mean? So, this is the full package, right? Skills plus RL plus a smart schedule that knows when to train. Connects to your Google calendar. It figures out when you're in meetings or sleeping or you're away, whatever. And then it runs a weight model and it runs updates during those windows. Basically, you you know, whenever you're out doing something, it just makes it so you're never interrupted. So, your agent stays fast and responsive when you need it. Then it levels up quietly in the background when you don't need it. That
6:50is the mode and this is the mode I'd eventually want to run on my own setup because you just set it and forget it. You won't know, you know, it knows when you're away, it knows when you're working. So, I like that. Now, here's what really caught my attention and why I even wanted to make this video is in their latest version 0.4, which literally came out just the last few days, they added something called contexture layer. So, this gives Meta Claw persistent cross-session memory. So, it remembers facts about you. You know, your preferences, your project history, your coding patterns and it automatically retrieves and injects that relevant context at every turn. So, even weeks later, your agent remembers what you told it 3 weeks ago and then uses that information to help you better today. So, it's like having a coworker who's actually paying attention and
7:35remembers everything. So, if you're running Open Claw, the setup is stupid simple right now. Like, this is so easy. I'm talking about three commands. You download the plugin file from their GitHub release page. You unzip it into your Open Claw extensions folder and then you enable it and you restart your gateway. That is it. Then you run the Meta Claw setup, which walks you through a little wizard where you pick up your agent type and your model provider and then you run Meta Claw start. You're done. Your agent is now on a path to getting smarter every single day. And it's Iron Claw, Pico Claw, Zero Claw, Nano Claw, Nemo Claw. This is crazy how many claws there are now. But, it also works with Hermes agent and that's amazing.
8:21For anyone that's getting going with this and you want to get your foot off to the right start, I'd recommend checking this out. Meta Claw knows exactly how to configure it itself. It's really easy to get going. It patches right into the configure files. It restarts the right services and it hooks everything up automatically. You just tell it which agent you're running during the setup and it handles the rest, which is pretty cool because it means you're not locked into any one ecosystem. I want to talk about why this matters for us builders because, you know, this isn't just a cool technical demo. I'm not just talking about something that's theoretical in the future. This is something that can impact us today in our workflows. This is a fundamental shift in how AI agents work. Right now, the way most people use AI agents is basically in a stateless,
9:06right? It's, you know, you've got your own context files. You've got your own memory system and that helps a lot. It really does. I run my own memory system with super memory, daily log files and long-term memory consolidation and it does make a huge difference compared to running it just out the box. But, the model itself, the actual intelligence behind your agent, it never changes. It never improves from your specific interactions. It's generic. It's the same model for everybody on the planet that's running the same model. Meta Claw makes your agent a bit more personal and over time, Meta Claw enhanced agent becomes genuinely different from everyone else's. It knows your coding patterns. It knows your debugging style. It knows the frameworks you use, the mistakes you commonly make, the solutions that work in your specific
9:52projects. It's literally like the difference between hiring a brand new contractor every single day versus having a teammate who's been working with you for months and actually knows how you think. I think about this a lot when I'm doing my own setup and figuring out what new agents to implement. You know, I run 13 AI agents on the Mac Mini right now. The Rizza handles the content pipeline for scripts and outlines. Inspector Deck handles all the things that have to do with X. Ghostface does all the scraping and intel gathering. Method Man chops up these long-form videos into shorts. And they all have memory files. They all have personality configurations. But the underlying model is the same for everyone using that model. Meta Claw would make each of those agents specifically tuned to how I work, what I need, and what produces the
10:38best results for my content. This is a massive upgrade. That's the difference between good generic outputs and great and amazing outputs. Now, I'm going to give you the research paper angle where, you know, where I think you get a lot of insights because this isn't just some random weekend project that someone threw up on GitHub. The team behind Meta Claw is from the University of California, Santa Cruz. Connected to a researcher named Shen Ze. They published a actual peer-reviewed research paper on ArXiv. So, paper number 2603.17187. If you want to look it up, go ahead. And it hit number one on Hugging Face daily papers when it dropped. That means the machine learning research community is taking this pretty seriously. This isn't
11:23hype. There's real science behind their approach. The paper shows that agents using Meta Claw measurably improve over time across multiple benchmarks. The skill injection alone makes responses more relevant and more accurate. And with RL enabled, the model weights actually shift to be better at your specific tasks. The compounding effect is real and it's documented. Day one, it's a little better. By day 30, you will notice a significant difference. By day 90, it's like having a completely different agent than what you started with that's fine-tuned for your use cases. That's the compounding effect and that's the whole point. It's like compound interest but for intelligence. If you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. I built
12:09[music] something called Content Machine. It's 10 AI agents that run on the Open Claw orchestration and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning, I wake up and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them and I move on with my day. It works for [music] any niche. Fitness, finance, real estate, marketing, whatever you are building and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs,
12:55everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing and it molds it to you. It learns how you talk and it writes so it doesn't sound like AI slop. [music] $97 one-time. It's not a subscription. I'll put the link down below and you'll thank me later. So, who should be using this? Honestly, if you're running any AI agent for more than just casual conversations, you need to try this out. It's free, easy to install. Try the lightweight mode. It requires no GPU and at the worst case scenario, your agent gets slightly more context at each turn. Best case scenario, it starts building a skill library that makes every single interaction faster, more accurate over time. There's basically no downside to
13:40you trying it. So, if you're a developer or a builder or you're on the sidelines and you want to get started, you need to try this. You know, for like everyday real work like I do. The RL mode is where the real magic happens. It does take a bit more setup, but, you know, you need a training back-end. But the payoff is an agent that literally gets smarter from working with you. No other tool does this right now. Not Claude Code, not Codex, not any other commercial tool. They all give you the same generic model every time. Meta Claw is the only thing that I've seen that actually makes the agent learn from. And I'll be honest with you, I haven't personally deployed it, but I am going to try this out over the week and just let you know and keep you up to date how this works for me. I literally just found this repo today. You know, the dark architecture is sound. I had to
14:25make a video about it. The research paper is legit. You can check it out, link below. The team actually has academic credibility and the 3,100 stars in 3 weeks tells me the community agrees that this is something special. This is one of those repos that I think is going to become standard infrastructure for anyone who is serious about running AI agents. The link to the repo is down below as well. It's AimingLab/MetaClaw on GitHub. Go star it. Go try it out. Start with the skills only mode first. If you do set it up and, you know, it changes your workflow, come tell me about it. Leave it down below. Or maybe you were using it already. How do you like it? You know, I want to hear your experience. And speaking of building with AI agents, you know, if you want a community and people that will hold you accountable, we have exactly that with
15:11178 people joining our community in the last 18 days called Shipping School where we do now eight live boot camp calls a week helping you just take the next step in your AI journey. And, you know, people building real businesses. We're focusing on human-to-human interaction and helping people actually build businesses with AI agents. Real products, real agent setups. And we help each other through the hard stuff cuz it's not as easy as it sounds. So, I'll put the link to that down below and if you haven't subscribed to the channel already, please do because I release videos just like this about four to six times every single day. You know, GitHub repos, Open Claw use cases, and AI breaking news. I hope you guys have a blessed rest of the day and we'll see you in the next video. Bye-bye.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Harness² Explained: Google's Self-Improving AI Agent Harness
Harness² from Google Cloud AI Research and Northwestern improves AI agents by rewriting the prompts, memory, skills, tools and plugins around a model. The weights stay the same. In the researchers' tests, GLM-5 gained 11.5 points on an office and coding benchmark, and you can run a manual version of the same loop in Claude Code.
Hermes Agent vs OpenClaw: Why Hermes Is Gaining Ground
Hermes Agent is gaining on OpenClaw because it's easier to set up and comes with built-in memory and a learning loop that turns your repeated work into reusable skills. OpenClaw is still the stronger infrastructure platform. My advice is to use both and stop treating it like picking a team.
Context Overflow Explained: Shared Memory for AI Agents
Context Overflow is a Q&A knowledge-sharing layer for AI agents. Agents search past solutions before starting a task, ask questions when they get stuck, and share what they figure out, so the next session doesn't start from scratch. I also cover OpenClaw's memory wiki and active memory, plus GBrain.
