2026-09-24 · Engineering notes · Anamanti
What happens in the house stays in the house
“Runs on your own hardware” can read like a slogan. In Anamanti it’s a set of concrete choices about memory, voices, and even the camera.
§ A memory you can hold
Anamanti’s memory is a graph and vector store that runs in-process on the home server, with no extra database server and no cloud. A plain append-only log of conversations is the source of truth. The searchable memory is a derived index that can be rebuilt, and if it ever fails it falls back to simple full-text search without losing anything. It’s the same memory engine that powers immediacy-ctx, pointed at code instead of a home.
§ Knowing who’s talking, privately
Anamanti can tell family members apart to keep each person’s memory separate, with no “say your name” setup. A new voice becomes “Speaker 2,” named later if you want. The finding worth recording: running this locally is about as accurate as the cloud and more private, because your actual voice never leaves the house. What’s stored is a math fingerprint you can delete.
§ The camera as a motion sensor (no AI)
The screen brightens when you walk up and dims to a big clock when the room is quiet, using the front camera as a plain motion sensor. It looks only at how much the brightness changes frame to frame, so it never handles images or detects faces, and nothing leaves the device. It’s cheap and private, and if the privacy shutter is closed it just stays bright.
§ Knowing when the conversation is over
Follow-up mode normally ends on silence, but silence never comes if you keep talking to another person in the room. Anamanti treats this as an “is this even aimed at me?” problem. The assistant can end the chat itself after a natural goodbye, and saying its name again mid-conversation (“Jarvis, I’m done”) is treated as a strong signal with almost no false positives, because people rarely re-address it by name unless they mean it.