Fusion Brain

A local-first, self-compounding knowledge vault for Obsidian — three ideas fused into one coherent system, plus a nightly self-healing audit.

• Karpathy LLM Wiki — raw sources flow into raw/, an LLM compiles them into clean interlinked pages in wiki/.
• Garry Tan Resolvers — every agent action consults C/RESOLVER.md first; no hardcoded paths, no context bloat.
• Mario Valle Reyes Main MD — C/HOME.md is a ~50-line front door every agent reads before anything else.

Why

A naive Obsidian + LLM setup loads 10k+ tokens on a simple query and files things wherever it feels like it. Fusion Brain caps base context at ~50 lines, enforces routing discipline, and composes fractally — sub-resolvers spin out per domain as the vault grows.

Targets:

• 70–100× token reduction vs. naive full-vault loading
• Zero misfilings over a 30-day window
• Graph View that stays meaningfully clustered instead of turning into a hairball

What's in this repo

Path	Role
C/HOME.md	The ~50-line vault map — every agent reads this first
C/RESOLVER.md	The mandatory routing decision tree
skills/SELF-HEALING.md	Nightly audit skill (orphans, broken links, drift, misfilings)
INSTALL.md	Full 9-step install guide

The repo layout mirrors the target install structure, so you can copy C/ and skills/ straight into your vault.

Quick start

# 1. Clone
git clone https://github.com/RegardV/Fusion-Brain.git
cd Fusion-Brain

# 2. Drop the core files into your vault
cp -r C skills /path/to/your/obsidian/vault/

# 3. Create the companion folders your vault still needs
cd /path/to/your/obsidian/vault
mkdir -p raw wiki sources

# 4. Initialize git so every compile becomes a rollback checkpoint
git init && git add . && git commit -m "Fusion Brain baseline"

Then follow INSTALL.md for the rest: enable Dataview, wire your existing skills through the resolver, run the first end-to-end test, schedule the nightly audit, and (optionally) go local-first with Ollama.

Prerequisites

• Obsidian with a vault you can restructure (or a fresh one)
• Git
• Claude Code (or any agent runtime) configured against the vault
• Optional: Ollama with a local model for offline, zero-leak operation

How it stays coherent

The SELF-HEALING.md skill runs nightly (or on /audit-vault) and produces a single audit report covering:

• Orphans — pages with zero incoming and outgoing links
• Broken [[wikilinks]]
• Contradictions across related pages (flagged for human review — never auto-resolved)
• Misfilings against the resolver's decision tree (with proposed git mv)
• Resolver traffic analysis — proposes new routes or sub-resolvers for hot paths
• File-count drift and graph health (emerging clusters, gap clusters)

Safe fixes auto-apply. Anything structural (moves, deletions, resolver edits, contradiction resolutions) is proposed for human approval only.

Credits

Fusion Brain v2.0 is a composition of three public ideas from builders who shaped the LLM-native knowledge stack:

• Andrej Karpathy — the LLM Wiki model that turns a pile of raw notes into an interlinked, compiled Graph View.
• Garry Tan — the resolver pattern for disciplined LLM routing: read the map first, never hardcode paths, no dark capabilities.
• Mario Valle Reyes — the Main MD / single-front-door convention that keeps base context around ~50 lines.

This repo fuses the three into one drop-in template plus a nightly audit. All errors of synthesis are mine.

License

MIT — use it, fork it, reshape it. If it helps your workflow, a link back is appreciated but not required.