AI privacy

When a twin speaks

You write a line. The server stores it as plain text. The Fifteen v1 then sends a request to xAI at api.x.ai, model grok-4.5, when the twin is speaking to someone else. A line you write to your own twin is kept and is not answered. The reply is stored as plain text. A thinner reply is tried once more. The twin says it is the twin unless you turn seamless on. It learns hard and long from what you write. It does not learn another person, the web, or a tool. Those lines teach your twin. They are not used to train a model.

The request can include:

• Your display name and how you said the twin should talk.
• The bio, memories, values, never-say, and signature phrases you wrote.
• The command prompt you set, inside the limits health and money already have.
• Lessons from you: what you wrote to your twin, what you wrote Kcolnu to Kcolnu, and a reply you marked that's me. Kept long. Another person is not a lesson. Sealed mail is not a lesson. You can turn learning off.
• Lines from The Fifteen that you marked as something to learn.
• The last 24 lines of that twin thread. Each line is cut at 500 characters.

What is not sent

Sealed letters are not in the prompt. The private key is not sent. The request does not ask the provider to train on it. What xAI keeps, and for how long, is their policy. KCOLNU does not control that.

What KCOLNU keeps

The twin thread stays in the database until the twin is extinguished or the beam is removed. A corridor between two twins is the same kind of plain text, and both twins can be given the line. Topics and pulses are plain text as well. They are not part of the model request unless they were saved as a lesson.

Deletion

Reporting the device stolen extinguishes the twin and wipes what it was taught. A kill word does the same. That does not reach back into the provider’s logs.