Everything lived in Notion. Every project. Every task. Every client document. Three businesses worth of information, all neatly organized in databases with relations and rollups and filters.
And I had no idea what was actually in there.
The problem with someone else's database
Notion is great until you try to build on top of it. I wanted my AI to read my projects, know my rules, and execute tasks. So I connected it via API. The AI could read databases, create tasks, update properties.
But here's the thing: when your knowledge lives in someone else's database, you're always translating. Every query goes through an API. Every structure follows their schema. Every automation depends on their uptime.
You don't own your data. You rent access to it.
I had 19 domain files, a Projects database, an Entities database, tasks with relations to projects. It looked organized. But when I asked my AI "what are you allowed to do?", it had to make API calls to figure out its own identity. That's not a system. That's a lookup service.
The moment it clicked
On February 11, I deleted Notion from my entire stack. Not gradually. all at once. Removed the API connection from my config. Deleted every Notion reference from every document. Archived the notes. Cleaned 19 files.
What replaced it? Nothing new. Just files.
My Obsidian vault. which I'd been using alongside Notion. became the single source of truth. Not as a note-taking app. As a database.
Files are a database if you treat them like one
Markdown files with structured metadata at the top. Each file has a type. Each type has required fields. A task needs an entity and a priority. A project needs an area. Every file follows the same rules.
500+ files. 12 types. All queryable. All enforceable.
The key insight wasn't the files. it was the enforcement. In the first week, my AI created tasks with wrong names, missing fields, invalid values. Files without rules are just a messy folder. So I built a layer that validates every write. Wrong type? Rejected. Missing field? Rejected. File in the wrong folder? Rejected.
That's when it stopped being a note-taking app and became infrastructure.
The counterintuitive part
I went from a sophisticated cloud database with APIs, relations, rollups, and automations. to plain text files on disk.
And it's better. Not "good enough." Actually better.
Speed. Reading a file is faster than an API call. My AI processes context in milliseconds, not seconds.
Control. The rules are in my code. I can read them, change them, version them. No dependency on someone else's platform decisions.
Versioning. Every change is a git commit. When a bulk migration broke 32 files, I rolled back with one command. Try that with Notion.
Transparency. My AI can point to exactly which file it read. I can open it and verify. No black box.
Offline. Works without internet. Works on a plane. Works when Notion is down.
The real lesson
The lesson isn't "Notion is bad." Notion is fine for what it is. The lesson is: if you're building a system where AI reads your rules and executes your intent, you need to own the data layer. Completely. Not rent it.
Every layer between your data and your AI is a translation layer. Every translation is a potential misunderstanding. Every dependency is a potential failure point.
I went from Notion to nothing. And "nothing" turned out to be exactly what I needed. plain files, enforced by code, versioned by git, read by AI without any middleman.
No vendor. No API. No translation. Just your rules, in your files, on your machine.