Why I built OS-Intelligence

AI without memory is a chatbot. That’s the line, and it’s been the through-line for everything I’ve built this year.

I run hundreds of conversations a week. Stakeholder calls, design reviews, half-thoughts on the train, three Slack threads that won’t resolve. By the time I sit down to make a decision, the things I need to remember are scattered across notes, transcripts, screenshots, my own head. The session that gave me the answer ended an hour ago and the answer left with it.

So I started with the obvious move. Better notes. Better tags. A project doc per workstream. Tried Notion, then a Slack channel for myself, then Obsidian. Each version solved part of the problem and created a new one. The system always sat outside the workflow, and so it always lost the thread.

What changed was Claude Code.

When I started using Claude Code daily, I noticed something. The model was capable enough to reason across a body of work, if you fed it the body of work. The blocker wasn’t intelligence. It was retrieval. The thing the substrate didn’t have was a way to capture, synthesise and serve back the context I’d already created.

So I started building. Not a tool, exactly. A system. Skills for the things I do every day. Discovery prep, PRDs, decision records, status updates. A memory layer underneath them. Sub-agents to review the output before I read it. I called it PM-OS and ran my entire job on it for three months.

After about six weeks, two things happened.

The first was that I started getting back time I hadn’t realised I was losing. The compounding kicked in. Every meeting fed the project memory. Every project memory fed the next meeting. I could ask “what did we decide about pricing in March” and get an answer with the receipts attached. It felt less like using a tool and more like working with a colleague who had been in every room I’d been in.

The second was that other PMs started asking how I’d built it.

That’s when I realised PM-OS was two things in a trench coat. Most of it was my workflow, and not very portable. But the layer underneath was different. Capture, synthesis, retrieval, the way memory was structured, the routing logic, the sub-agent fan-out. That layer was generalisable. It became OS-Intelligence.

OS-Intelligence ships the reasoning layer in the open. Apache 2.0, one-command install, every entry is a markdown file in your repo. It’s not a chatbot, and it’s not a notes app. It’s the substrate underneath those things. The thing the model needs to actually be useful across the whole job.

The bet I’m making is simple. AI-native work is going to need a memory layer. The companies that win will be the ones that figure this out early. The PMs who win will be the ones who’ve already built a system that compounds.

That’s what OS-Intelligence is. The follow-ons will be about how I built it. The choices I made, what I tried first, what I’d do differently. Each one a decision record, in the open.