Knowledge flow
Knowledge flow
Raw captures become compound intelligence through a pipeline that runs automatically. You never trigger it. You just work - and the knowledge compounds.
The flow
Capture → Seal → Compound → Synthesis → Sub-vaults → Knowledge
Capture. While your session is open, fai writes JSONL files continuously - AI transcripts, file diffs, git events, terminal commands. Zero AI calls. Zero cost. Each capture is also indexed for semantic recall.
Seal. The CaptureVault compounds on a time-gated cadence, promoting pending
captures to kept. This fires an OnKeep event that routes instructions to
downstream knowledge vaults.
Compound. Each vault owns its own compound schedule. When the threshold fires, the vault's compound strategy runs — synthesis agents query captures via semantic recall rather than reading raw files.
Synthesis. Four synthesis agents run independently — one per sub-vault (Context, Pattern, Decision, Petnames). Each agent uses recall-filtered context relevant to its domain and produces updated vault knowledge.
Kept. The synthesized knowledge commits to the kept branch. This is the
stable, distilled record of what your workbench has learned from all sealed sessions.
Sub-vaults. The four sub-vaults (ContextVault, PatternVault, DecisionVault, PetnamesVault) each hold their synthesized knowledge as structured files.
Knowledge. The KnowledgeVault aggregates all four sub-vaults into:
FAI-PROJECT.md- everything this project knowsFAI-PERSONAL.md- everything about how you work
Every AI you connect reads these on startup.
What triggers synthesis
Synthesis runs automatically when enough sessions have sealed. You configure
this threshold when you run fai synthesize:
fai synthesize
The interactive setup asks which synthesis provider to use (Ollama, LM Studio, OpenRouter, Groq, OpenAI, claude-code, or capture-only) and how frequently to synthesize. After that, it runs on its own.
The result
By session 10, your AI starts a session already knowing:
- Your architecture decisions from the last 9 sessions
- The naming conventions you've been using
- The patterns you apply consistently
- What you said you'd work on next
No re-explaining. No context pasting. Just: continue.
Pure mechanical observation
Everything you do is written to disk as structured JSONL. Git activity, file changes, terminal output, AI transcripts. No model calls, no cost. Runs continuously. This is the foundation - nothing can compound without this.
Patterns, decisions, context
AI agents query captures via semantic recall and synthesise meaning into the appropriate vault types. Each sub-vault runs its own compound on its own cadence.
The moment your vault becomes your brain
The full cascade: captures promoted → OnKeep routes to sub-vaults → each compounds independently → knowledge vault aggregated. Runs automatically on each vault's own cadence. After this runs, every connected AI gets your full accumulated knowledge injected on startup.