Why imprnt?
I am the guy who spends two weeks buying a bike for a fifteen-minute commute. Any bike would do, and I know it, but I crave the conscious decision, which turns every purchase into an investigation nightmare: a week of trail bikes and tire types no commuter needs, a week of test rides, then a bike way over budget that fits the commute perfectly if you don’t count the money.
It was like that for ages, with everything, and I know people who think alike. Information became accessible, and mediocre stopped being acceptable. Even when it costs time.
The good news: it’s 2026, and you can afford a personal assistant. A real one, that knows your life and takes over the boring part of being you. I built mine.
It started with Claude. My research moved into chat, with a friend who knows something about everything and has no memory. Every session I briefed a brilliant stranger who happened to have read the entire internet.
Three lines of the same briefing before the one line I came to ask. Every chat.
- The bike I spend two weeks buying a bike for a fifteen-minute commute.
- It started with Claude Every new chat I briefed a brilliant stranger from zero.
- PAI, for someone else PAI remembered me, then became more machine than my life needed.
- The token bill I had become afraid of my own assistant.
- The rule A machine able to launch spaceships ended up reading spam letters.
- The bible Eight rules, so the model thinks once and searching is free.
- Memory is plain text Open any note, fix one line, and take the folder anywhere.
- My promise to you What you don't need, it doesn't keep, and everything's opt-in.
PAI was perfect, for someone else
Then I found PAI, I don’t even remember where. It did the one thing I wanted most: it remembered, and it gave my goals a written structure called TELOS. For a while it was exactly right.
Version 5 is where we parted ways. It grew machinery I never asked for: a scheduled dashboard, a Telegram bot I had no use for, a framework of ideal goals and criteria that was more machine than my life needed. Every piece makes sense for its author, who writes about AI for a living. Mine is a different job.
Then came work, with no flat subscription. Tokens used to be almost free for me, and suddenly every question had a price tag. I caught myself counting every token, skipping things I could figure out otherwise. I had become afraid of my own assistant.
An assistant you are afraid to use is not an assistant.
The rule that fixed it
Around that time my colleague Forud Ghafouri showed me how he works with AI agents, and I felt blessed. He named the trap I was sitting in: people treat it like magic that should therefore do everything. Don’t get me wrong, AI can do almost everything. That doesn’t mean we have to ask it to.
His rule goes the other way: the model is a peer you bring in for judgment, and your own code is the servant that does the work. Do it backwards and a machine able to launch spaceships ends up reading spam letters.
You would not pay a sharp colleague to re-read your whole inbox every morning to find the three emails that matter. You would have them write the filter once. After that the filter runs every morning, for free, forever.
- Has judgment
- Costs real money
- You spend it once, to build the tool
- Does the mechanical work
- Costs nothing to run
- Gives the same answer every time
That was my whole PAI problem in one line. So I wrote my assistant a list of rules, brief and to the point, the way a bible should be:
The bible
- Never do the same work twice. Build the tool, run the tool.
- Leave dumb work to dumb code.
- Never lock my knowledge behind you.
- You remember things so I don’t have to.
- No order, no action.
- What I don’t need, you don’t keep.
- Update the fact, keep the history.
- My notes have my voice, not yours.
That list was the start of imprnt. Token use went down, results went up, because I was finally using the model for what it is good at.
This is the setup that did my German taxes, and my case was far from a textbook one. A week to get my head around taxation here, the bureaucracy done, the return received, thousands saved on tax advisors.
And taxes are just one folder. The same housekeeping runs my documents, my purchases and my hobbies, and what I figure out stays solved instead of evaporating with the browser tabs, the way the bike research did.
Your memory is plain text
Most AI memory tools store what they know about you as numbers, built for the model to match and unreadable by a person. Your memory is reachable only through the tool that wrote it.
imprnt keeps your memory as plain text notes in one folder on your computer. Search is plain math on your machine, so nothing depends on someone else’s server. Claude Code and Gemini CLI both run it today.
- Open it like any text file and read it.
- A wrong fact is one line. Edit it, done.
- Switch assistants: the folder comes with you.
- Rows of numbers. Nothing to read.
- A wrong fact has no line to find.
- Switch assistants: export first, if the tool allows.
The comparison walks the field tool by tool, with credit where another tool fits better.
My promise to you
One rule I will not break, because breaking it is what pushed me out of PAI. Line 6 is my promise to you: what you don’t need, it doesn’t keep. No surprise dashboard, no background daemon you did not ask for. I have been shipping software for 13 years and I don’t ship broken things.
What ships is essentially a folder of markdown files. Your plain Claude stays plain, everything else is opt-in, and anything you add comes out with one command.
The over-researcher in me is alive and well, he just runs on a better platform now. The all-knowing nerd next to me already holds my sock size, my detergent dose and my beer tiers, so the next obsession costs an evening, not two weeks. The bike decision these days would take fifteen minutes.
Start here. The base is small on purpose, and the plugins are there when you need one.
MIT licensed. © 2026 Aleksandr Bogdanov.