How Noa works
Noa is the AI companion in MindMate. This page explains what she does, what she works with and where her limits are – as transparently as the EU AI Act requires and as we want it ourselves.
A conversation, not counselling
Noa listens, sums up what you said, then asks one question – one, not five. She uses techniques from cognitive behavioural therapy (examining thoughts, spotting patterns), acceptance and commitment therapy (values, distance from thoughts) and motivational interviewing (your own why). She does not diagnose, judge or push.
What Noa remembers
Not the conversation but the summary she writes at the end: title, themes, one to three insights, a next step. Next time she knows the last three of those, your first name, your goals and your latest mood. All of it is visible in the app and can be deleted.
Speak or type
In voice mode your phone recognises what you say and Noa answers in one of two voices – “warm” or “calm”. The text of the answer appears alongside. In text mode you type.
Safety
Every message is checked for signs of acute distress before the answer. When the protocol fires, Noa changes tone, stays with you and the app immediately shows the crisis line and the emergency number. Noa alerts nobody and cannot avert an emergency – she tells you so.
Technology
Noa's answers come from large language models (Anthropic, OpenAI) via the intermediary OpenRouter, with zero data retention and EU data centres preferred for text. Voice output uses an OpenAI audio model without zero data retention – details in the privacy policy. An answer usually starts after about one and a half seconds.
Limits
- Noa can be wrong. Language models sound confident even when they are not.
- Noa is no substitute for therapy – she is the bridge to it. That is why the therapist directory lives in the same app.
- Noa is labelled as an AI and will always say so when asked.