Active · Local-first AI · Phase 4

MAX3

A local-first, voice-first AI resident — a single Python process that lives on one box, not a service that visits.

The MAX3 Brain panel showing the ontology navigator: a force-directed graph of 6 disciplines, 37 subjects and 265 topics, with a node-select sidebar.
The Brain panel — the ontology MAX3 reasons over: 6 disciplines, 37 subjects and 265 topics.
PythonDuckDBGraph memoryForward-ForwardVoice-firstLocal-firstFlagship

The concept: a resident, not a visitor

Most assistants are a service you call. MAX3 lives on one box in the house: his memory, his learning and his voice all run there, and he stewards the household — the network and the off-grid power — the way a resident would.

Why local-first

  • The data never leaves the box. Memory, retrieval and learning run on hardware the household owns.
  • Every action is reversible. He acts only on PIN-gated, reversible, audited changes.
  • Every claim is checked. He checks a claim against a council of models before trusting it, and writes new skills into an approval queue rather than running them.

How he works

  • Memory — a pure-DuckDB graph spine, with full-text and vector retrieval fused into one ranking.
  • Learning — a Forward-Forward Variable-Vector learner grows meaning in the edges: predict, prove, adjust, never rote.
  • Voice — a wake word, streaming Whisper and sentence-chunked Piper speech, so you just talk to him and can interrupt him mid-sentence.
  • Curiosity — a durable ingestion bus reads arXiv, GBIF, World Bank and a dozen more open sources, and a curiosity loop notices its own gaps and goes looking.
  • Judgement — he reasons with emotion, a six-lens morality framework and an explicit epistemic state.

Screenshots

Follow the collective

New research, projects and field notes as they publish. Following happens on LinkedIn — we never see an address, and there is nothing to unsubscribe from.

Or get it by email

An occasional summary of new research, each with a link to the full piece.