Mojo

Mojo 🔮

A local-first tacit-knowledge reference layer that turns Claude/Codex sessions, commits, notes, and decisions into scoped advisory context for the next task.

The Problem

Every Claude Code session produces valuable decisions, corrections, and domain rules:

• "No — in our environment we use B instead of A, because…"
• "This framework has an N+1 trap here, always prefetch…"
• "Don't revert that, it was failing because of the cache key, not the query."

And then the session ends, the context window is gone, and the next session starts from scratch. The knowledge is in the transcript, but nothing consumes it.

The Promise

Mojo captures work evidence through local hooks and backfill commands (both **Claude Code and Codex** sessions), distills durable tacit knowledge with a two-stage LLM pipeline (cheap filter → strong structurer), and keeps quality separate from applicability scope.

Three things make the loop sustainable:

1. Zero marginal cost by default. The pipeline runs on headless claude -p (your subscription), not an API key. codex exec and the Anthropic API remain available as backends.
2. Knowledge lives as markdown in an Obsidian vault. One claim = one note; reviewing = editing frontmatter. The knowledge is visible where you already read and think, not locked in SQLite.
3. Generalization is evidence, not filing. Every claim starts project-scoped; re-observing it in a second project records evidence, and only cross-project support suggests promotion. A counterexample sharpens does_not_apply_when instead of deleting the claim.

Your workflow stays human-controlled: Mojo surfaces what matters, why it was selected, where it applies, and where it should not be generalized.

Reference Layer, Not Auto-Injection

Mojo's default project workflow is attach/detach:

mojo attach --project ~/code/my-service
mojo refresh --project ~/code/my-service --task "debug auth cache"
mojo status --projec