Self-learning AI: an assistant that learns from the people who use it.

A large language model is trained once, on a fixed snapshot of text, and then deployed. From that moment its knowledge and habits are frozen. Everything that looks like learning afterwards — long system prompts, retrieval, “memory” features — is scaffolding bolted on around a model that has not changed.

A self-learning AI closes that loop. Interactions feed back into the system and alter how it answers next time. Done honestly, that means three separate things, each with a different speed and a different risk profile:

• Personal learning — remembering what you told it, for you alone. Fast, low-risk, and it should always be visible and reversible.
• Collective learning — promoting what many people independently taught into knowledge every conversation can draw on. Powerful, and dangerous without a gate.
• Weight-level learning — actually retraining the model on verified examples so the improvement lives in the checkpoint, not in a prompt.

Most products that market themselves as learning AIs do the first, quietly, and neither of the others. conw.ai does all three, and the reason it can is that each layer is guarded differently.

Instant: your private memory

Say it in a sentence — call me Tom, keep it short, glimble means a tiny useful shortcut — and it is saved before the reply comes back. The next answer that relies on it says so, with a Forget button right there. No settings panel, no syntax. The exact phrasings that work are documented in How to teach Conw something that sticks.

Days: guarded shared knowledge

A taught word or fact can go further than your account, but only through a gate. It must be taught by several independent people, days apart. It is screened for safety and for anything personal, then reworded so it is never anyone’s literal sentence. It is quizzed from angles nobody taught it. What passes is released to a small canary slice of conversations and watched; only then does it become shared knowledge — and it can still be pulled. Every decision, kept or refused, is written to the public ledger at conw.ai/learning.

Nights: weight-level retraining

Eligible examples wait in a guarded queue. When there is enough clean material and the serving machine is idle, training runs in the background and yields the moment live chat needs the hardware. The output is a candidate checkpoint, not an automatic update: it has to load and pass its checks before anything live changes. The model behind an interface stays replaceable, so memory, safety rules and user ownership survive every checkpoint swap.

Anyone can claim their model learns from users. The claim is only worth something if you can see what it refused. conw.ai publishes a running rejection ledger: what people proposed, what was promoted into shared knowledge, and what was thrown out — with the reason attached. Not enough spread. Screened out. Conflicting accounts. Failed the quiz. Demoted after going live.

Most of what it is taught never makes it, and that is the point. A model that can refuse you is a model worth teaching. The ledger is unauthenticated and updates continuously: see everything it was taught, and everything it threw away.

“Memory” has become a checkbox on every assistant. Here is what the checkbox usually means, next to what conw.ai does.

Open a chat and say any of these in plain words:

• “call me Tom” — a preference about you
• “reply in under 30 words” — a preference about how it answers
• “EOD = end of day” — your own shorthand
• “no, glimble is a small workaround” — correct a definition it got wrong

The next reply that uses it will tell you so. If you would rather keep the whole thing on your own hardware, Conway Entity is the same self-learning loop running locally on your machine, with nothing leaving it.