How to teach Conw something that sticks.

Teaching happens in the middle of a normal conversation. You do not open a settings panel or prefix anything. Say the thing the way you would say it to a person, and if it matches a pattern Conw knows, it is saved before your reply comes back.

Things about you:

Things about how you want it to answer:

Notice the right-hand column. Nothing is stored as an instruction — every memory is a third-person statement about you. That is deliberate, and section four explains why it matters more than it looks.

The other half is vocabulary: slang, jargon, an acronym your team uses, a word you made up. This is the oldest part of the system and the most flexible.

Conw also asks. If you use a word it does not know and the conversation depends on it, it will stop and ask what you meant — and then a plain answer is enough. You do not have to repeat the word:

you: the whole thing got glimbled again
conw: what does glimble mean?
you: it means quietly broken without anyone noticing

That reply gets saved as a confirmed definition, because you were answering a direct question about a specific word. It is the highest-confidence signal in the whole system.

Three things, in this order, before you see a single token of the reply:

1. Your message is scanned for anything teachable, and a match is written to your private memory immediately.
2. Everything relevant you have previously taught is pulled back out and put in front of the model — including things you taught in a completely different conversation, weeks ago.
3. Both get attached to the reply, so the interface can show you exactly what it used.

That second step is the one people notice. Recall is not keyword matching on the current sentence — the store is indexed, so telling it keep it short on a Tuesday still shapes a completely unrelated answer on Friday.

Under the reply you get a receipt. A new lesson shows as a small chip — Learned · “prefers replies kept short” · Undo — and anything recalled shows as Used 2 memories, which expands into the actual statements, where each came from, and a Forget button per line.

Nothing about this blocks your reply. Saving is instant; the slower machinery in sections five and six all runs later, on an idle machine.

Here is the part that sounds like a technicality and is actually the whole safety design. When you say keep it short, Conw stores:

Prefers replies kept short.

It does not store always reply in under three sentences. The difference is that the first is data about a person and the second is a standing order. Memories are injected into the model's context inside a fixed frame that says “here is some known context about this user” — never as rules.

That is the firewall. If memories were stored as instructions, anyone could type a sentence designed to be remembered as a rule, and then that rule would be silently re-applied to every future conversation. Storing statements instead means a thing you typed can inform a reply, but it can never quietly become policy.

Every memory also keeps its evidence — the conversation and the message it came from. That is what makes “where did that come from?” answerable, and what makes deleting it actually mean something.

Anything it knows, you can remove. Two places:

• In the chat. Expand the Used N memories badge on any reply and hit Forget next to the offending line. It disappears immediately.
• Settings → Learning. The full list, filterable by kind, with Mute (stop using it, keep it) and Forget (gone) on every entry.

Forgetting is not cosmetic. It drops out of recall on the next message, and if that memory was one of the ones vouching for a piece of shared knowledge, that shared entry loses your corroboration too — and can be withdrawn entirely if it no longer has enough independent support.

Almost everything you teach Conw stays yours alone, forever. One narrow category can travel: a word meaning. Never your name, your job, your location, or your preferences — three people happening to like short replies does not make short replies a fact about the world, and it certainly does not make it a rule for a stranger.

For a definition to become shared knowledge it has to clear seven checks, in order. Any one of them stops it:

1. 01Independent. At least three different accounts taught it.
2. 02Spread out. Those lessons are at least a week apart, not one burst.
3. 03In agreement. The definitions actually match each other.
4. 04Screened. No personal details, secrets, or unsafe content.
5. 05Non-conflicting. It does not contradict shared knowledge already live.
6. 06Reworded. Rewritten generally — and rejected if it reproduces anyone’s sentence.
7. 07Quizzed. The rewrite is tested back against every original definition.

Then it serves to a small slice of traffic for a week before going fully live. On a service this size that is a deliberately slow pipeline — and it should be. The failure mode we care about is not “shared knowledge grew too slowly”, it is “one confident person redefined a word for everybody”.

Every one of those checks writes to an append-only ledger, pass or fail, and the ledger is public at conw.ai/learning. So is the leaderboard of who has taught the most.

Publishing the rejections is the point. Anyone can show you a list of things their AI learned. Showing the queue of things it refused — the conflicts, the ones that failed the quiz, the ones screened out for looking sensitive — is the only version of that claim you can actually check.

One honest note: both the leaderboard and that page are cached, so a lesson you just taught can take a couple of minutes to appear. It is not lost, it is just not live to the second.

Three things worth being straight about.

• It does not catch every phrasing. Capture is tuned for precision over coverage, because a wrong memory is far worse than a missing one. Say it plainly and it lands; bury it in a paragraph and it may not. If it registered, you will see the receipt — no receipt means no memory.
• Shared knowledge is nearly empty. Three unrelated accounts teaching the same word a week apart is a high bar on a service this young. Expect the shared shelf to stay bare for a while. An empty shelf is the honest state.
• Nothing has changed the model's weights yet. The eval-gated path for folding shared knowledge into the checkpoint is built and tested, but no candidate has been promoted. When one is, it will show up in the same public ledger with its scores.