Skip to main content
Reflections is a voice-first AI agent that retrieves and reasons over a personal knowledge graph. Users talk to their “reflection” — an LLM backed by chunked documents, entities, and temporal facts ingested from uploaded sources.

Architecture

System overview, execution planes, boundary rules, and package structure.

Decisions

31 architecture decision records documenting core technical choices.

​
Key concepts

  • Two-plane architecture — realtime inference is separated from background learning for safety and reliability.
  • Temporal knowledge graph — facts are append-only with temporal validity, supersession, and gated promotion.
  • Voice-first interaction — conversations are orchestrated through a managed voice provider with server-tool callbacks.
  • Eval-gated learning — candidate facts must pass evaluation before becoming active truth.