The video is in Finnish.

OpenClaw is an open source project for building a personal AI agent — one that runs 24/7, responds to your messages, performs scheduled tasks and connects to your messaging channels. It's an impressive system: a long-running Node.js process, reconnect logic for ten messaging platforms, a heartbeat mechanism, cron schedulers and sandbox security layers.

But I haven't built an OpenClaw agent. I've been maintaining my own agent for a couple of months — a much simpler one — and in this video I show how you can build one too.

Persistent Memory

The most important feature of an agent is persistent, long-term memory. For my agent, this memory consists of files: focus.md describes what I'm currently working on, preferences.md defines my technology choices and .cursorrules tells the agent how to work with me.

OpenClaw does the same: it has AGENTS.md, SOUL.md, USER.md and HEARTBEAT.md. These files are injected into the LLM's system prompt — exactly the same mechanism. The difference is that OpenClaw has much more long-term memory and the agent modifies its own memories autonomously.

Security and Control

An OpenClaw agent autonomously modifies its own behaviour, can activate at any time and has integrations with messaging platforms, email and GitHub. That's why it includes a 7-layer tool policy pipeline, sandbox isolation and an approval workflow for critical actions.

My system has one control loop: me. The agent does nothing without me. The attack surface is small because the agent never acts alone.

What Kind of Agent Do You Need?

I don't need a creature that reads my messages and does things on my behalf. But less can be enough for many — and costs stay under control because API calls only happen when I ask.

If this sparked any thoughts, get in touch!