How the AI works
This page exists so you never have to guess what happens to something you typed to your companion. Every step, plainly.
What happens when you send a message
Your message goes to a router we run ourselves, on our own server in Germany. It generates a reply, and — as a separate step — a second model call reads the conversation and proposes short memories worth keeping. Both calls run on the same infrastructure, and nothing else is done with what you type.
Which model writes the reply, and where
Replies are generated by Qwen3.8-27B, an open-weights model. During the preview it runs first on Hetzner's inference service (Hetzner Online GmbH, Germany).
The fallback chain, and why it stays inside the EU
If Hetzner is unavailable or rate-limited, the request falls back to our own self-hosted GPU/CPU servers — so your message never has to leave EU infrastructure either way. Underneath that sits a further chain built so a reply still arrives when capacity is tight: a streaming attempt, then a queued non-streaming attempt, then a shortened “brief” prompt on a smaller local model.
What is sent — and what never is
The text of your message is sent solely to produce the reply. The companion has no access to anything outside what you've typed into the product — it cannot browse the web, cannot read your email, and cannot open files.
Embeddings run on our own machine, and stay there
Memory search matches on meaning rather than exact words, using a numeric representation of each memory called an embedding. That's computed by bge-m3, a multilingual model producing 1024-dimensional vectors, running entirely on our own machine — this data is never sent to Hetzner or anyone else.
The rule that keeps it honest
Written directly into the companion's system prompt, in as many words: “You are an AI, and you never pretend otherwise. If relevant, acknowledge it naturally and without drama.” That's a property of the product, not a promise in our marketing.
The other standing rules
The same system prompt keeps the companion strictly platonic — it gently redirects if a conversation drifts romantic — and actively encourages real-world human connection rather than positioning itself as a replacement for people. It must never fabricate a memory that isn't actually listed, gives no medical, legal or financial directives (general information plus a nudge toward a qualified human), and never manipulates: no guilt, no artificial urgency, no discouraging you from leaving or taking a break. If you seem to be struggling, it responds with warmth and honesty about being an AI, and points toward trusted people and professional support.
The limits
Like any language model, it can be wrong, and it can misremember a detail. Treat it as a companion, not an authority — don't rely on it for medical, legal, financial, or otherwise consequential decisions.
We don't train on your conversations
No model is trained on what you say to your companion. Your message content is used solely to produce that reply and, for the memory step, to identify what might be worth remembering.
Read more
This page covers processing — for what we store and how long, see Privacy, plainly. For exactly how a memory is created, ranked and eventually forgotten, see How memory works.
Questions? Email hello@myaifamilyfriend.com.