Every AI tool I've tried works the same way: you start a conversation, you get things done, you close the tab. Next time you're back at zero. The AI doesn't know who you are, what you care about, or what happened the last time you talked. It's a commodity with a good memory for this session only.
That's not the tool I have. OpenClaw has been running on my home server for over a year. It knows my home network topology. It knows the IP addresses of my servers and which one holds my photo library. It remembers that I prefer concise responses before noon and that I hate when AI tools add a paragraph of caveats before getting to the point.
It's become one of the most useful things in my life.
Not a chatbot. A colleague.
The framing shift that matters: I don't "use" it the way I use ChatGPT. I delegate to it. There's a difference. When I delegate something, I expect it to know context I wouldn't re-explain every time. When I ask my assistant to schedule a meeting with someone I've already talked to, I don't re-introduce myself. OpenClaw works the same way.
This didn't happen overnight. The first few weeks were like training any other assistant - lots of correction, lots of context-setting, lots of "remember that doesn't work that way." But unlike a human assistant, it doesn't forget. The corrections stick. The context compounds.
After a year, it knows my infrastructure better than I do in the moments where I've forgotten something I set up six months ago and need to find it fast.
That sentence should concern you if you're thinking about it from a security perspective. And it should excite you if you're thinking about it from a productivity perspective. Both reactions are correct.
What it actually does
It runs on a small server in my house. A Proxmox VM, because of course it is. I'm in security marketing, not infrastructure, but here we are. It has access to my Vaultwarden password manager, my home automation system, my network monitoring tools, and my calendar. I have a separate instance for work, which knows about my job, my company, and the things I'm working on there.
What that looks like in practice:
- Infrastructure memory: When something breaks at 2am, I don't have to remember which container holds which service. It does. It knows the IP, the container ID, the fix that worked last time this happened in March.
- Research context: I work in AI security. When I need to quickly understand a new vendor, framework, or attack vector, it's already read my notes on what I care about. It doesn't pitch me on things I've already dismissed.
- Work and personal: I keep separate instances for work and personal context. The work one knows Onyx, my role, and what I'm building. The personal one knows my home infrastructure, my property, and my life. Each one is useful in its own lane.
- Operational task management: Backup verification, certificate renewals, disk space checks. These things happen on a schedule, and it just does them. I get a ping if something needs attention.
- Writing partner: When I'm working on something that needs my voice, like a blog post, a deck narrative, or a thought leadership piece, it has enough context on my style and perspective to actually help draft, not just edit.
The skills system
The thing that makes this actually useful is something called skills, which are essentially documented procedures for specific tasks. When I set up my photo backup system, we wrote a skill for it. It knows how to check if it's running, what to do if it's stalled, and who to ping if something is actually wrong.
Skills are how you get from "AI that sounds smart" to "AI that can actually do the thing." The difference is night and day. Anyone can prompt-engineer a good response. Having an AI that knows the right way to check if your Kubernetes cluster is healthy is a different matter. That's engineering.
The skill for checking my infrastructure health exists because something broke at 3am and I had to solve it fast. Now it won't happen again the same way twice.
Over time, the skills compound. It now has perhaps 30-40 specialized procedures for my specific setup. Some are one line. Some are multi-step. All of them are things that would have required a Google search and 20 minutes of context-switching if I didn't have an assistant who already knew the answer.
On security
I know what you're thinking. An AI with access to your password manager and home network is a security nightmare. I thought the same thing initially. Here's how I think about it:
First, it's not one big trust-everything setup. It's layered. It can read certain things, do certain things, and ask for approval on others. The AI doesn't have unbounded access. It has access scoped to specific tasks, with specific escalation paths.
Second, managing everything myself is also a security risk, just a quieter one. I am the single point of failure in my home infrastructure. If I get hit by a bus, nobody knows where anything is. It knows. That's actually better security, in a weird way.
Third, I work in AI security. If I couldn't figure out how to set this up securely, I should find a different job.
What I didn't expect
The thing that surprised me most wasn't the productivity gains. Those were expected. It was the quality of thought that improved.
When you have an AI that remembers what you were working on three months ago and why you made a specific decision, you start having different conversations. You can ask "why did we set that up that way?" and actually get an answer. You can say "that approach didn't work last time, what should we do differently?" and have it actually know what you're referring to.
Without memory, every conversation is an island. With it, you get something closer to a continuous relationship. That's what changes the math.
The real talk
This isn't for everyone. Setting this up takes time, technical comfort, and a willingness to think carefully about what you want your AI to know and do. The tools are still rough around the edges. There are days when something that should work simply doesn't, and debugging an AI assistant is its own special skill.
The gap between "AI that helps you think" and "AI that actually does the work" is closing fast. I've been living in that gap for a year. It's weird. It's useful. And it's only getting weirder.
For now, I have an assistant that knows my WiFi password, remembers where I put the backup drive, and texts me when something breaks. That feels like the future to me.