Introducing Nychorium: private AI without the asterisks

We built Nychorium because we got tired of every AI product treating our conversations like training data. So we did the opposite. Here is what that actually means, what is live today, and what we are still figuring out.

The state of the AI industry, briefly

Almost every mainstream AI service does the same three things to your chats. They log them. They tie them to your account and your real identity. And, in many cases, they use them to train the next version of the model. The privacy pages say it more politely, but that is the shape of the deal.

That means the moments you would never put in a Google search, questions about your health, your relationship, your finances, a career change you have not told anyone about yet, all of it is sitting in a database somewhere with your name on it. Maybe behind good security. Maybe not. Either way, you are not the only person who can read it.

That is the part that bothered us most. The trade was never explained clearly. You signed up for an assistant and got a permanent record.

Our take

Nychorium is a private AI chat service. The hard rules are simple:

• No chat history on our servers. Your conversations live in your browser. Clear your browser, they are gone. We do not have a copy.
• Anonymized at the model boundary. When your message reaches the AI, it is stripped of anything that identifies you. The model sees the question, not the person.
• Never used for training. Not by us. Not by the model providers we route through. We pick partners that contractually agree to this and we drop the ones that drift.
• Token counts, not content. The only thing we keep about your usage is how many tokens you sent and received, because that is what billing needs. The actual words are not on our disk.
• No ads. No data brokers. No tracking pixels. The page you are reading does not have a single analytics script that follows you around.

What is live right now

Frontier-grade chat that does not moralize

You get an AI that thinks before it answers, handles long context, reads files and images, and runs web searches when it needs current info. It will tell you that it disagrees with you. It will give you a direct answer. It will not lecture you for asking a normal question.

Voice chat

Talk to it like a person. Voice in, voice out, with the same privacy posture as text. Nothing about the audio is stored on our end. The transcription happens, the model responds, the audio is gone.

Memory, in beta

Some people want their AI to know them. Their name, the projects they are working on, the way they like to be talked to. We get it. We also know memory is the single feature that has the biggest privacy implications, so we built ours carefully.

• Off by default. If you never turn it on, nothing is ever saved.
• End-to-end encrypted. Memory entries are encrypted in your browser with a key only you hold. Our servers see ciphertext. We genuinely cannot read what you have saved.
• Recovery key in your hands. When you turn memory on, we generate a key, show it once, and you save it. You paste it on any other device you want to link.
• Delete means destroyed. Wipe a memory and it is gone. We do not soft-delete or archive.

It is in beta because the model behavior around memory is genuinely hard. How much should it volunteer? When should it stay quiet? How do we avoid the failure mode where it brings up something personal at the wrong moment? We are iterating on it in public.

Why we do not name the models we use

A few reasons, all of them honest.

We use several different models behind the scenes. Different jobs get different tools. A quick reply does not need the same brain as a deep research question. We route to whichever frontier-grade model gives the best answer for what you asked.

We also swap models when better ones come out. Naming a specific model in our marketing would either become a lie three months from now or lock us into something inferior. We would rather quietly upgrade you to the best available.

And finally, the choice of which providers we trust is a privacy decision, not just a quality decision. We will not route your data through a provider that retains or trains on it, no matter how impressive the benchmarks look. That filter is part of what you are paying for.

Tiers

Our philosophy

Privacy first, always. When two design choices conflict, the more private one wins. That is the rule, and we hold ourselves to it even when it makes the product harder to build.

But privacy on its own is not enough. There are plenty of privacy-respecting products that nobody uses because they are miserable. We want Nychorium to be the AI you actually reach for, because it is good, not because it is virtuous.

That balance is the whole game. Memory is a good example. The most private version is no memory at all. The most useful version is total recall. We landed on optional, encrypted, deletable, in your control. It is not perfect. It is the honest middle, and we will keep refining it.

What is next

This product has a long way to go. We have a list. Better mobile experience. Smarter memory recall. More languages. More polish on voice. More options for how you organize your conversations locally. We will ship them as they get good enough to ship.

If something feels off, tell us. If something feels right, tell us that too. We read everything.