Building My Own OpenClaw on a Mac Mini
There’s a moment in every side project where you realize you’ve accidentally learned more than you intended.
For me, it started with OpenClaw — Peter Steinberger’s outrageously good self-hosted AI assistant. If you haven’t seen it, go look. 191k stars on GitHub. Messaging integration across everything. 100+ skills. A real community building real things with it. It’s the project that made me go “holy shit, I want to understand how all of this works.”
So I started building.
Taking things apart to understand them
I’ve always learned by building. Not by building products — by building the thing that helps me understand how the thing works. Personal AI systems are the most interesting thing happening in software right now, and OpenClaw is the best example of one. So I’m pulling it apart, learning how the pieces fit, and putting my own version together along the way.
The plan is to use as much of OpenClaw’s work as possible and hopefully contribute back. I’d rather be part of Peter’s community than off in a corner reinventing wheels.
What I’ve been tinkering with
Six sessions of “just setting up infrastructure” taught me more than I expected:
- Inngest for durable workflows — I wanted to understand event-driven architecture, and holy shit does it click when you build it yourself
- Qdrant for hybrid search — dense + sparse + BM25. I didn’t know what any of that meant a month ago
- Redis for state and pub/sub — turns out this is where a lot of the magic lives
- Tailscale + Caddy for networking — making services talk to each other securely
- Video pipelines — download → transcript → summarize, all wired through event chains
None of this is better than what OpenClaw provides. It’s just me learning how the pieces fit together by building them from scratch. You don’t really understand plumbing until you’ve flooded your own basement. 😅
The memory rabbit hole
The part that’s fascinated me most is memory. How does an AI system remember things? Not just “stuff the last 5 messages into the prompt” — real memory.
I’ve been designing a 4-layer system:
- Session recall — what happened in recent conversations
- Playbook — patterns extracted from repeated work
- Timeline — narrative history of what was built and why
- Soul — identity, values, personality that persists across everything
Is this overengineered for a learning project? Absolutely. But that’s the fun part. I get to think about how memory should work without worrying about shipping it to anyone. No deadlines. No users. Just me and a question I find genuinely interesting.
Standing on Peter’s shoulders
I wouldn’t be building any of this without OpenClaw. The idea of a self-hosted agent-as-OS with messaging as the primary interface — that’s Peter Steinberger’s vision, and it’s the right one. His Lex Fridman interview is a great place to start if you want to understand where this is all going.
My hope is that as I learn, I can give back. Document the things I figure out. Maybe contribute a skill or two. At minimum, write about the journey honestly enough that someone else who’s curious can learn from my mistakes.
That’s all this is. A guy learning in public, grateful for the open source work that made it possible. 🔥
This is the first in a series about building a personal AI system from scratch. Next: AT Protocol as the bedrock for personal AI.