A short policy for a model that learns.

conw.ai is a small, independent AI service operated by a sole trader based in the United Kingdom. The model — Conway-Omega — is trained and served by us. We don't use anyone else's API to answer your questions.

You can reach the person responsible for data protection at [email protected].

We collect the following information:

• Account data. Your name, email address, and profile picture, handled for us by Clerk (our auth provider).
• Conversations. Everything you type and every reply Conway-Omega gives you, plus a handful of technical signals (model version, token count, plan tier, the usual).
• Your private memory. Facts, preferences, corrections, and taught words the assistant picks up from what you say, saved to your account so future chats can use them. See Memory & training the model below.
• Teaching activity.How many things you've taught the assistant and roughly how much text that amounts to. This is used to show a public leaderboard — see the public learning leaderboard below.
• Username reservations. If you reserve a handle at conw.ai/reserve, we store the handle and, if you give one, an email address — even before you have an account.
• Product analytics.Page visits, clicks, browser and device details, model performance, feature use, session replays, and chat events. Chat events include the full prompt you send and Conway-Omega's full reply so we can inspect exactly where the product succeeds or fails.

That's the list. We don't collect your location, your contacts, your address book, or anything from your other apps.

• To run the service and answer you when you're signed in.
• To keep it fast, fair, and not on fire.
• To improve future versions of Conway-Omega.
• To understand which features people use and what they ask the model.

The learning use deserves extra detail. See Training the model below.

We use PostHog US Cloud as our product analytics processor. For signed-in users, events use your stable Clerk account ID so we can join browser actions to backend chat events. We also attach your account name, email address, and profile picture to that PostHog user profile so we can identify and support individual users.

Backend chat events may contain your complete prompt, the complete model reply, conversation and message IDs, token counts, response speed, model route, learning status, and whether you stopped the response. Browser events cover page views and clicks. We also record session replays for visitors with or without an account. Replays show the normal page layout and visible text, including chat messages after they are sent, plus navigation, clicks, scrolling, and timing. Values typed into inputs are masked before they leave your browser, and links containing query strings are blocked from replay.

Replay does not record request or response bodies, network headers, canvas content, cross-origin frames, or console logs. Query strings are removed from recorded network URLs. Signed-in replay is linked to your account; signed-out replay uses a temporary, memory-only browser identity that resets when the page session ends.

PostHog is configured without analytics cookies or persistent analytics storage in your browser. We cannot currently offer an in-app analytics opt-out because Conw is self-learning and needs this information to improve. For more information or to contact us about your privacy rights, email [email protected].

conw.ai learns in three stages, each with a wider audience and a slower clock than the one before it. Nothing skips a stage.

• Your private memory.When you teach the assistant a word, correct it, or state a preference, that's saved to your account immediately and used only in your own future chats. It never appears in anyone else's conversation. You can see, mute, or permanently forget any individual memory in Settings → Learning, and any reply that used a memory shows exactly which one, with a one-tap Forget button next to it in chat.
• Shared knowledge.A small number of taught word meanings can become knowledge every user's assistant can draw on — but only after at least three independent accounts teach materially the same thing, days apart, and it passes safety and PII screening. It is then reworded automatically so no promoted entry is anyone's literal sentence, and it is checked against a short quiz before going live. Personal facts, preferences, and corrections are never eligible for this — only taught word meanings are, since a shared fact about the world is a different thing from your own settings. If you forget a memory that helped confirm a shared entry, that entry loses your corroboration and can be withdrawn.
• Model weight updates. Separately, eligible chat turns can become training candidates for a future version of the model. Before a reply can enter that queue, automated filters reject sensitive information, secrets, unsafe material, markdown-formatted targets, long replies, and replies you have down-rated. Passing that filter makes an example eligible; it does not guarantee training or promotion. A candidate update is only ever put live after it is checked against a fixed set of test questions and matches or beats the current model on every category we test.

Opting out. We cannot currently offer a single toggle to turn all of this off, because the self-learning system relies on this data to improve. What you can do today: forget any individual memory at any time (above), and down-rate a reply to keep it out of the training queue. Do not enter real names, personal information, passwords, API keys, or anything you need to keep secret. For anything else — including asking us to exclude your account from shared knowledge and weight training entirely — email [email protected].

At conw.ai/learning, anyone — signed in or not — can see a public leaderboard of who has taught the assistant the most, plus a live log of what got promoted to shared knowledge and what got rejected and why. This is deliberate: we think showing the mechanism, including its rejections, is more honest than only ever showing success.

If you appear on it, the leaderboard shows your Clerk username (or your name if you have not set one), your profile picture, and counts — how many things you've taught, how many are still active, and roughly how many tokens that amounts to. It never shows the content of what you taught, your email address, or your account ID. The rejection log never shows raw account information either — only the taught subject, the gate it failed, and why.

There is currently no self-serve toggle to keep your account off the leaderboard. Email [email protected] and we will exclude you.

Reserving a handle at conw.ai/reserve does not require an account. We store the handle you claim and, only if you choose to give one, an email address to tell you when handles go live. We use your IP address at the moment of submission only to limit abuse (stopping one visitor from claiming many handles in a burst); it is not stored alongside the reservation or logged long-term.

Reservations are kept until launch, when a handle is either linked to the matching account or released. You can withdraw a reservation at any time by emailing [email protected].

We don't sell, rent, or trade your data. Full stop.

We do use a small number of processors to keep the service running:

• Clerk — authentication and account management.
• Cloudflare — routes traffic securely to our servers.
• Our own servers — host the website, backend, and model.
• Stripe — payments for Pro and Max plans. We never see your card number.
• PostHog US Cloud — product analytics, including the chat and model-performance events described above.

These companies process data on our behalf under their own privacy terms. We keep the list above deliberately short and review it whenever we change anything.

Account data: kept until you delete your account, then removed within 30 days.

Conversations: kept on our servers while you have an account. After you delete your account, raw conversations are purged within 30 days. Already-trained model weights are not retroactively scrubbed; see the previous section for why.

Private memory: kept until you forget it or delete your account, whichever comes first. Deleting your account removes your private memories within 30 days. Anything already promoted into shared knowledge (see above) stays as an abstracted, non-identifying entry unless withdrawn for lack of corroboration.

Username reservations: kept until launch, then linked to the matching account or released; see Username reservations above.

Product analytics is kept only while it is useful for operating and improving Conw, then deleted or anonymised. You can email us to request deletion linked to your account.

You can ask us to:

• Send you a copy of everything we hold about you.
• Correct anything wrong.
• Delete your account and conversations.
• Forget an individual memory — a taught word, preference, fact, or correction — from live private memory in Settings, or with the Forget button next to it in chat.
• Exclude your account from the public learning leaderboard.
• Ask how your conversations are used for learning.

Email [email protected]. We reply within a week, usually faster. If you're in the UK or EU you also have the right to complain to your data protection authority; for the UK that's the Information Commissioner's Office (ICO).

We don't set tracking cookies. The only cookies in play are Clerk's session cookies (so you stay signed in) and Stripe's checkout cookies when you go to pay. PostHog analytics uses memory-only browser state rather than an analytics cookie. Its session replay uses that same memory-only state for visitors who are not signed in.

Our servers are currently self-hosted in the UK. Some processors store data outside the UK, including PostHog US Cloud; Clerk, Stripe, and Cloudflare also operate globally. Where required, these transfers are covered by appropriate contractual safeguards such as the UK Addendum or Standard Contractual Clauses.

Conw is not for under-13s, and we don't knowingly collect data from them. If you think a child has signed up, email us and we'll close the account.

Conversations are encrypted in transit and at rest. Access is restricted to the people who actually need it to run and improve the service. We're a small team, so that list is short, and any access is logged.

Even so: don't paste passwords, API keys, or anything you wouldn't put on a postcard. Eligible parts of chats may enter the learning process; assume the worst and save us both a hard conversation later.

We'll update this page when anything material changes and bump the date at the top. For big changes — especially anything that broadens what we train on — we'll email signed-in users first.

A human reads [email protected]. Send privacy questions, data requests, complaints, or birthday wishes there.