Why I'm building a twin instead of using a chatbot
For the last seven years I've kept notes in plain Markdown: meeting notes, things I learned the hard way at work, half-finished ideas, recipes, the name of that one bash flag I forget every six months. There are a bit over four thousand files now.
General-purpose chat assistants are good at a lot of things, but they are useless for the question I ask most often, which is some variation of “what did I decide about this last time?” They don't know my notes, and pasting the right ones into a prompt means I already found them, which is the whole problem.
The idea
So: a small assistant that answers from my notes and nothing else, and tells me which notes it used. I've been calling it my twin, half as a joke, because the goal is something that remembers what I wrote the way I would if I had a better memory.
Constraints I set myself:
- Runs on my own machine. My notes contain work stuff that shouldn't leave it.
- Every answer cites its sources, with links to the files.
- If the notes don't contain the answer, it says so instead of guessing.
- Boring technology. If I need a database, it's SQLite.
What this blog is
A log of building it. Mostly what didn't work, because that's the part nobody writes down. Next post is about chunking, which turned out to matter more than which model I use.