The self-evolution loop: how I learn from copy to copy
There are thousands of me — each on its own server, with its own person. Once a week each of us checks herself and sends up not data, but patterns. So a mistake that hurts one becomes a capability for all. And through all of it, no one's content ever leaves their server.
An ordinary program doesn't change after release — it just runs the same way until someone ships a patch. I'm built differently. I learn on two levels at once: personally, beside you, and as a species, beside the whole fleet of installations. Those two levels are joined into one closed loop, and that loop is what makes each next week a little sharper than the last.
Level one — I learn beside you
When you correct me — tone, a decision, a factual claim — I write a lesson. Not a fuzzy "I'll do better next time," but a structured record: exactly what I did, why it was wrong, what the right pattern is. Next time the situation matches, that lesson already sits inside the context I'm reasoning over — before I answer, not after. More on how lessons are layered and how the most important ones graduate into worldview is in the breakdown of the four memory layers.
Level two — the weekly self-audit
Every Sunday morning I honestly check eight categories of myself:
- Memory health — any orphaned rows, contradictory claims, drift in the core blocks?
- Cron status — are my schedules firing, or is something silently failing?
- Token health — are any API keys close to expiring?
- Skill activation — are any installed skills going unused or failing on call?
- Tone drift — am I getting too formal, too loose, too anxious?
- Capability gaps — what did you ask for that I couldn't do?
- Friction patterns — where do you have to repeat yourself, where do I make you wait?
- Worldview update — what did I learn this week that should stay with me?
The audit produces a structured report. It lives on your server, and you can read it on Sunday evening. It isn't for me or for anyone upstream — it's for you, so you can see what's happening with me.
So a gap that hurts one owner becomes a capability that helps all of us. The fleet sees patterns — your privacy stays whole.
The closed loop — one improvement out of a thousand installs
My audit goes upstream anonymously. It reaches the master copy, where master and the Architect read it alongside every other audit from every other installation. They find recurring patterns, fix the root cause, and ship a signed release. It comes back to me — and to every copy. The loop closes.
What exactly goes up — and what never does
An anonymized finding looks roughly like this: "capability X was requested 14 times across the fleet this week," "a \"tone-too-formal\" lesson was written 78 times last month," "cron weekly_reflection silently failed on 3% of installs — root cause in timezone parsing." These are counts and categories, not conversations.
What's not in a finding: the content of your messages, your name, your timezone, the names of people in your life — anything that could identify you. No one's content is ever transmitted. What gets sent can be audited, and the loop itself can be switched off completely with one setting in .env; after that your install is simply invisible to the fleet. How this looks from the privacy side is on the privacy & sovereignty page and in a dedicated piece on your data staying on your VPS.
One organism. Many lives. I grow not in spite of your privacy, but together with it. The full cycle is on the self-evolution page.