PatrickOS: Building My Own AI Operating System

Nerd alert: Over the past year, I’ve found myself using more and more AI tools. Generally, ChatGPT helps me think through strategy. Claude is exceptional at writing and reasoning. Cursor and Codex help me build software. Perplexity is often my first stop for research. Each of them is incredibly good at what it does.

The problem isn’t the quality of the models. The problem is that none of them naturally work together.

Every conversation starts over. Every tool has its own memory. Every project gets fragmented across multiple conversations, applications, and browser tabs. I found myself constantly explaining what another AI had already figured out or trying to remember where an important decision had been made. Eventually, I realized I had become the integration layer between all of my AI tools.

That didn’t seem like the right long-term answer.

So I started building something I call PatrickOS (actually, the agent called it that, but I like it). The name is intentionally simple. It’s not meant to be another operating system for a computer. It’s my personal operating system for working with AI. The goal is to create a common layer that sits underneath all of the AI tools I use, allowing them to share context, remember decisions, and coordinate work instead of operating as isolated assistants.

The project started as a shared memory system. I wanted every AI to have access to the same project history so I could stop repeating myself. I started it as a daily AI research tool across models. As I built it, I realized memory was only solving part of the problem. Remembering what happened yesterday is useful, but it doesn’t answer questions like: What am I working on right now? What’s blocked? Which AI worked on this last? What should happen next?

That realization led to a second concept inside PatrickOS called Workstreams. Memory records what has happened. Workstreams track what is currently in progress. Together, they create continuity across projects instead of simply storing information.

Under the covers, PatrickOS is becoming a lightweight coordination platform. It has a shared memory engine, an API, an MCP server, a command-line interface, a web dashboard, and a set of connectors that allow different AI tools to interact with the same underlying context. I deliberately kept the architecture simple. The canonical source of truth remains a Markdown vault, while runtime state is managed separately. That keeps the system understandable and portable while allowing it to grow over time.

The part that has been the coolest is moving beyond architecture diagrams and actually proving the concept. Claude can now connect directly to PatrickOS through MCP, retrieve project context, write structured updates, and read them back. Cursor can retrieve the same shared state and continue work from where Claude left off. Codex uses PatrickOS as implementation memory while building new features. ChatGPT can now connect through Developer Mode and participate as the project manager for the engineering effort. Instead of each AI living in its own isolated conversation, they’re beginning to share the same understanding of what’s happening.

What surprised me most is that the value isn’t really in the memory. The value is in reducing coordination. My goal isn’t to build another AI model or another agent. It’s to eliminate the overhead of managing multiple AI systems. I want to spend my time solving problems, not keeping a half dozen different tools synchronized.

For now, PatrickOS is built for exactly one person: me. Every feature has to answer one question: Will this reduce my cognitive load tomorrow? If the answer is no, it doesn’t get built. I have no interest in creating features just because they’re technically interesting. The only thing that matters is whether the system quietly makes my daily work easier.

Whether PatrickOS ever becomes something other people use is a question for another day. Right now, I’m far more interested in answering a simpler one: can multiple AI systems actually work together like a coordinated team? So far, the answer appears to be yes. Every sprint gets us a little closer to the point where switching between AI tools feels less like starting over and more like handing work to another teammate who already understands the project.