Skip to main content
PROJECT B.R.A.I.N. exists because useful long-term knowledge needs more than a model session. A conversation can end, a model can change, and a provider can be unavailable. The project keeps the information that matters in a durable system the user can inspect, correct, back up, and move.

​
Persistent memory and knowledge

B.R.A.I.N. stores structured cognitive records, conversation events, document metadata, and processing history in PostgreSQL. It preserves original and normalized document artifacts separately. Search indexes and embeddings improve retrieval, but they are derived and rebuildable; they are never the only copy of knowledge.

​
Provenance and correction

Every accepted teaching can retain its source, authority, confidence, context, and history. When a user corrects something, B.R.A.I.N. creates a successor and marks the earlier record as superseded. The earlier state remains inspectable, so the system can explain both what it currently believes and why that changed.

​
Provider independence

Gemini currently provides replaceable capabilities such as Live interaction and embeddings. Gemini’s own knowledge, session state, and output are not B.R.A.I.N.’s canonical knowledge. The Conversation Gateway records the durable interaction and routes provider work without transferring ownership of memory or identity.

​
Implemented today

  • Private, authenticated text teaching, recall, explanation, and correction flows.
  • Durable learning lineage with active and superseded history.
  • A bounded background worker and restart-safe migrations.
  • Gemini Live text and raw-PCM interaction through the shared Conversation Gateway.
  • Ingestion of extractable PDF, DOCX, Markdown, and plain-text documents with versioning, portable originals, model-tagged derived embeddings, and grounded citations.
  • Backup and isolated-restore evidence for application data and document artifacts.

​
Deferred capabilities

Phase 7 daily-use qualification is in progress. Its technical prerequisites passed, but the required three-day genuine-use window is not complete. Home embodiment, sensor ingestion, a world model, autonomous schema changes, model training, a local resident model, and optional GPU acceleration are later work. They are not implied by the accepted Phase 6 state.