Why imprnt?
In 2026, I wanted to buy a bike to ride to work. A normal person would have walked into the nearest shop, pointed at one, and taken it home. I started researching instead, and two weeks later I knew more about gears, suspension travel, and types of racing bikes than anyone shopping for a commuter will ever need. I did buy a bike in the end. It just took two weeks of homework to get there.
If you recognize even a little of yourself in that story, keep reading. That is how I do everything. A hobby, a purchase, a decision - I have to understand a thing inside out before I act. The research used to run through Google. Then it moved to Claude, and Claude felt like a very smart friend who knows something about everything.
But that friend had no memory. Every chat started from zero. I would type out my setup, my constraints, the decision we had made the week before, and by the next session all of it was gone. Every time, I was briefing a brilliant stranger who happened to have read the entire internet.
The same briefing, typed again at the start of every chat.
- The bike I cannot buy anything without understanding it inside out first.
- Ask Claude Brilliant answers, zero memory. Every chat met a stranger.
- PAI Finally a tool that remembered me, with real depth.
- Version 5 It grew machinery I never asked for and could not remove.
- The token bill At work, metered, the bloat ate the budget I needed to work.
- A colleague's rule Treat the model as a peer. The script is the servant.
- imprnt Spend the model once to build the tool. Run the tool free, forever.
PAI was perfect, for someone else
Then I found PAI, I do not even remember where. It did the one thing I wanted most: it remembered. It carried context across sessions, it gave my goals and values a written structure it called TELOS, and it had the depth I had been missing. For a while it was exactly right.
Version 5 is where we parted ways. It grew machinery I never asked for: a Pulse dashboard that ran on its own schedule, a Telegram bot wired in permanently when I had no use for one, a whole framework of ideal goals and ideal criteria that was more machine than my life needed. Every piece made sense for its author, who writes about AI for a living. A tool that helps him write all day is exactly the right tool for him to build. Mine is a different job.
Then I took it to work, where there is no flat subscription. Work runs on a metered budget, where every token is real money. On the first load, against an almost empty account, that machinery burned through tokens before I had typed anything useful.
The problem had changed shape. The model used to forget too fast. Now the parts I never asked for were eating the budget I needed for the actual work, and I could not pull them out.
The rule that fixed it
Around that time a colleague reframed the whole thing for me. He was ruthless about tokens, and his rule was simple: the model is a peer you bring in for judgment, and the software you write is the servant that does the work.
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
- Runs a thousand times, identically
That was my whole PAI problem in one line. I had been treating a brilliant peer like a servant, paying full price for the same mechanical work over and over, so everything got slow and expensive. imprnt flips that around. The model reads a source and files a clean note once, with judgment. Plain local code does the searching and the checking a thousand times after, for free.
What you keep
The other half of the argument shows up the day you want to leave. Most AI memory tools keep what they learn about you in a store built for the machine: your words converted into long lists of numbers, so a model can match them by meaning. A person cannot read that store. A wrong fact about you has no line you can find and fix, and when you move to another tool, those numbers stay behind.
imprnt stores your memory as plain text notes in one folder on your computer. You can open any note and read it. A wrong fact is one line in one file, fixed in seconds. The search is plain arithmetic on your machine, so your memory never depends on someone else’s server, and the day you switch assistants the folder comes along. Claude Code and Gemini CLI both run it today.
If you want the field walked tool by tool, with credit where another tool fits better, the comparison does exactly that.
What imprnt will never do
There is one rule I will not break, because breaking it is exactly what pushed me out of PAI. imprnt will never be shaped around only my problems, and it will never hand you a tool you did not ask for. The default setup ships with almost nothing. Your plain Claude stays plain. Everything is opt-in, and anything you add you can remove with a single command. You take what serves you and leave the rest.
If that is the assistant you have been wanting, start here, or browse the plugins and add only the ones you need.