Memory and reasoning for your app, over an HTTP API.
A model is the mouth. It has no memory of the person it is talking to, so every app wrapping one rebuilds context from scratch each session, badly, out of whatever fits in a prompt. This is the other half: a graph that accumulates what someone actually said, narrows it to what matters for the message in front of it, and hands the model only that.
You bring three things — a model key, an embedding key, and a Neo4j instance. The brain brings the part that is hard: what to remember, what to surface, and what to leave out.
You bring the keys #
Your Anthropic key runs the voice. Your Gemini key runs the embeddings. Your Neo4j holds the graph. All of it is encrypted with per-tenant KMS envelopes and decrypted only at the moment of use. We never pay for your inference, your storage, or your recall — which is why one flat price works no matter how much you think, and why your data never sits in somebody else's database.
What you get #
POST /v1/filter
— a turn: retrieve, notice, narrow, speak.POST /v1/ingest
— write back what the turn learned.- Your own Neo4j — an Aura free instance is enough to start.
- Facts, standing instructions, open threads, entities and relations, moods, conditions.
- Every fact carries where it came from and when it was last confirmed.
No card, no trial clock. You bring the keys and the graph, so your usage costs us nothing — which is why we can leave this open until it's worth charging for.
Your key is issued straight away and shown to you once.
Before you start #
The whole integration manual is public — read it first and decide if it fits. talk to her instead · health · test console