Before there was an agent, before there was a name, before any of the architecture that came later, there was a folder full of markdown files.

I’d been using Obsidian for a while by then. Nothing revolutionary. Just a vault where I dumped everything I knew about my own setup. Infrastructure docs, debugging notes, code patterns I kept reaching for, that sort of thing. The kind of personal knowledge base that every engineer starts with good intentions and then either abandons or lets grow into something unwieldy. Mine was somewhere in between.

The vault existed because my setup is, to put it kindly, involved. I’ve got a homelab running virtualisation, home automation, a media stack, and various other services. At work I’m doing infrastructure and platform engineering at a startup, so there’s cloud platforms, Kubernetes, Terraform, CI/CD pipelines, and about fifteen other things that each have their own collection of gotchas and tribal knowledge.

That’s a lot of context to hold in your head. It’s also a lot of context to repeatedly explain to an AI.

The re-explaining problem

This is what got me. I’d been working with AI tools at work for months. Not just using them, building AI-powered products. I could see how capable the models were. The reasoning was impressive. The code generation was genuinely useful. But every single conversation started from zero.

“I’m using GKE with ArgoCD and Terraform. Our staging project is called X. We use TypeScript with strict mode, Prisma for the database, tRPC internally. I prefer types over interfaces. British English, please.”

Every. Single. Time.

It’s like having a brilliant colleague who gets a complete memory wipe every morning. You wouldn’t tolerate that from a human. You’d write an onboarding doc. You’d give them access to the wiki. You’d say “read this first, then we’ll talk.”

So that’s what I did.

Feeding the vault

The Obsidian vault started growing with a new purpose. I wasn’t just documenting things for my own reference anymore. I was documenting them for a future AI that would need to understand my world.

Infrastructure topology. Deployment pipelines. Code conventions. The fact that I will die on the hill of British English in codebases. My preference for early returns and composition over inheritance. The specific way our staging environment differs from production. Which monitoring alerts are actually important and which ones fire every Tuesday and can be ignored.

It grew organically. Every time I caught myself re-explaining something to Claude or ChatGPT, I’d think “right, that needs to go in the vault.” Every debugging session that took too long because the AI didn’t know our setup, that became a new note. Personal preferences, technical decisions, even the reasoning behind why we chose one approach over another.

None of it was fancy. No special tooling, no custom plugins, no elaborate tagging system. Just markdown files in folders, written the way I’d write notes to a colleague who was about to join the team.

The gap

Here’s what I kept bumping into at work. We were building AI products, genuinely useful ones, and the models were getting better every few months. But the gap between what an AI could do and what it actually knew about your specific situation was enormous. The intelligence was there. The knowledge wasn’t.

Every AI tool I used was smart but ignorant. Capable of incredible reasoning about problems it understood, completely useless about the specifics of my infrastructure, my codebase, my preferences. It’s the difference between hiring someone with twenty years of experience and hiring someone with twenty years of experience who has never seen your codebase.

The vault was my first attempt at bridging that gap. Not a framework. Not a product. Just structured knowledge in a format that could be fed into a context window.

What if the AI already knew?

That was the question that wouldn’t leave me alone. What if, instead of starting every conversation with ten minutes of context-setting, the AI just… already knew? Knew your stack. Knew your preferences. Knew that the last three times you tried this particular approach it didn’t work, and here’s why.

I didn’t have a plan for how to make that work properly. No architecture diagram, no roadmap. Just a growing collection of markdown files and an increasingly stubborn conviction that this was the right problem to solve.

The vault would eventually become the seed data for something much bigger. But in November 2025, it was just a bloke with too many notes and a question he couldn’t stop thinking about.

- Alex