Teach it once. See what stuck.

Correct a reply, rate it, or explain a new word. Confirmed meanings enter your private memory immediately, so cognira can use them in the same conversation. There is no settings panel and no special syntax — say it the way you would say it to a person.

A preference, a fact, or a word is written down before the first token of the reply comes back. If it registered, you will see a receipt. No receipt means no memory.

Sensitive, unsafe, badly formatted, and down-rated replies stay out. Only short, useful examples can enter the reviewed learning queue. A wrong memory is worse than a missing one, so capture is tuned for precision over coverage.

Almost everything you teach stays yours alone. The one narrow exception is a word meaning — never your name, job, location, or preferences.

For a definition to help everyone it has to clear seven checks: independent people, days apart, in agreement, screened, non-conflicting, reworded so it is nobody’s literal sentence, and quizzed. Any one of them stops it.

Background updates wait for enough clean examples and an idle server. A candidate must pass those checks before it can replace anything live. Confirmed teaching still works immediately from private memory — weight updates never block your reply.

Every rejection is public too, at /learning. Showing the queue of things it refused is the only version of “we learn in the open” you can actually check.