Bonfire
Architecture

The Knowledge Layer

A local-first LanceDB-backed knowledge graph that accumulates patterns across runs. Fork the repo and the knowledge layer is empty.

What it is

A local-first, LanceDB-backed semantic store that accumulates patterns across pipeline runs. The Knowledge Layer is Bonfire's answer to the "agents don't remember yesterday's lessons" problem that every other framework leaves to the application layer.

Where it lives

FilePurpose
knowledge/backend.pyLanceDB backend adapter
knowledge/ingest.py (214 lines)Ingest pipeline
knowledge/chunker.pyContent chunking
knowledge/hasher.pyContent deduplication
knowledge/embeddings.pyPluggable embedder (Ollama + mock + real options)
knowledge/scanner.pyRepo-wide content scan
knowledge/memory.pyIn-memory backend for tests and dev
knowledge/consumer.pyEvent-driven ingest from the pipeline event bus
knowledge/backfill.pyBackfill from session history

What it stores

  • Session logs (per-run)
  • Scout reports
  • Synthesis briefs
  • Wizard findings (both APPROVE and CHANGES_REQUESTED)
  • XP events
  • Pattern anomalies

Fork = empty brain

This is intentional. When you fork the framework or clone it fresh, your Knowledge Layer is empty. Over weeks of use, it grows rich with your patterns — your code style, your test shapes, your Wizard-review tendencies.

The knowledge graph is yours. It cannot be copied by forking the repo. It can only be earned by using the tool.

Three-phase rollout

The Knowledge Layer rolls out in phases gated by founder evaluation, not a release calendar.

PhaseWhatGate
0 — Dogfood (current)Knowledge Layer runs in the founder's personal forge. Real functionality evaluated on real workflows.Founder decides the experience is real, not hoped-for.
1 — Pro tierEach paying user gets an individual local vector DB.Phase 0 convinces the founder the layer earns its keep.
2 — Enterprise tierShared knowledge across a team (LanceDB Cloud backed).Phase 1 proves stable at scale.

The website will never promise a milestone on a date. If this page says "Phase 1" with no date, that's the honest state.

Collective intelligence (Phase 2 design target)

A shared knowledge graph across an enterprise team is designed but not implemented in v1. The API layer exists in docs/architecture-design.md as a spec. It rolls out after the Pro individual layer proves stable.

On this page