Build Your Brain in Obsidian
12 min readThe local-first version of your company brain. Full control over your data, a knowledge graph Claude can query at runtime, and the three-layer architecture that turns a note vault into a real AI context engine.
Obsidian is the other half of the brain. Everything is local, nothing touches the cloud, and the vault becomes a knowledge graph Claude queries at runtime.
The fastest way: let Claude build it for you
Open Claude Code in the folder where you want your vault and paste this. It scaffolds the whole structure, interviews you, and writes your CLAUDE.md and memory files in one go.
You're setting up my brain · my second brain · as a local Obsidian vault. Build it, then seed it from a short interview.
1. Create this vault structure (ask me where, default ~/obsidian/brain/):
00-home/ (index.md, top-of-mind.md) · atlas/ (projects.md, vault-structure.md) · inbox/ · knowledge/ · sessions/ · voice-notes/
2. Interview me one question at a time: what my business does and how it earns, my clients, my daily stack, how I communicate (with 2-3 lines of my real writing), and my top 3 priorities right now.
3. From my answers, write:
- a CLAUDE.md at the vault root · under 150 lines, a teaching document, not a config file
- a memory/ directory with decisions.md, clients.md, projects.md, patterns.md, open-loops.md, each starting with a one-line header explaining what belongs in it
4. Any note you create uses prose-as-title · every title is a falsifiable claim, not a category.
Print the CLAUDE.md in full when you're done and tell me how to point Claude at this vault.Everything below explains what that prompt builds and why · the three layers, the structure, and prose-as-title. Read it so you can shape the vault to how you actually think.
Notion or Obsidian · which for what
Both are excellent. They solve slightly different problems.
Notion is better when:
- You're building something collaborative · team or client
- You want a database structure for tracking things
- You want AI baked into the workspace
- You want drag-and-drop over files
Obsidian is better when:
- You want everything local and private · nothing in the cloud
- You want a knowledge graph with bidirectional links between notes
- You're building a deep personal knowledge base with thousands of notes
- You want maximum compatibility with Claude's file-reading
Many operators run both: Obsidian for personal knowledge and thinking, Notion for clients and projects. That's a valid setup.
The three-layer architecture
This is what makes Obsidian a real AI context engine, not just a note app.
┌─────────────────────────────────────────┐
│ Layer 3: Ingestion Pipeline │
│ video/audio → structured knowledge │
├─────────────────────────────────────────┤
│ Layer 2: Knowledge Graph │
│ Obsidian vault + MCP bridge │
├─────────────────────────────────────────┤
│ Layer 1: Session Memory │
│ CLAUDE.md + auto-memory directory │
└─────────────────────────────────────────┘Layer 1 · Session Memory (CLAUDE.md). What Claude reads at the start of every session: your identity, your current projects, your conventions. Keep it under 150 lines. It's a teaching document, not a config file.
Layer 2 · Knowledge Graph (the vault). Your searchable long-term memory. Claude accesses it via MCP. The vault is a persistence layer Claude can query at runtime, not just at session start.
Layer 3 · Ingestion Pipeline. How new knowledge gets in · transcripts, voice memos, recordings, articles, processed into structured notes automatically. (Covered in full in the next lesson.)
They compound. Skip one layer and the others degrade.
Vault structure
Structure matters enormously · an unstructured dump degrades retrieval fast. Use these levels so Claude finds the right depth on the first pass:
obsidian-vault/
├── 00-home/ ← maps of content · table of contents for your brain
│ ├── index.md
│ └── top-of-mind.md
├── atlas/ ← structural overview of everything
│ ├── projects.md
│ └── vault-structure.md
├── inbox/ ← unprocessed captures land here first
├── knowledge/ ← curated, processed knowledge
├── sessions/ ← session transcripts and call notes
└── voice-notes/ ← transcribed voice capturesThe highest-leverage move: prose-as-title
Name notes as claims, not categories. Every note title is a falsifiable claim. If you can't evaluate the claim from the title alone, rewrite it.
❌ memory-systems.md ✅ memory graphs beat giant memory files.md
❌ sales-notes.md ✅ warm outbound converts 3x better than cold when the sequence mirrors the buyer's language.md
❌ pricing.md ✅ anchoring the retainer to avoided FTE cost closes faster than ROI projections.md
When Claude searches your vault, the note title alone tells it whether something is relevant before reading a single line. Prose titles make semantic search dramatically more accurate.
The same applies to links. Make wiki-links read as sentences:
we learned that [[memory graphs beat giant memory files]]
when we [[benchmark retrieval like search infrastructure]]The graph becomes self-documenting. Retrieval becomes navigation.
Connecting Obsidian to Claude
Two MCPs give Claude structured and full access to your vault. Add them to Claude Code with claude mcp add · the same way you wired up Notion and GitHub in the Foundations module:
# Full read/write access to your Obsidian vault
claude mcp add obsidian -- npx -y obsidian-mcp
# Structured queries over notes by path, tag, and frontmatter
claude mcp add qmd -- npx -y @tobilu/qmd mcpPoint each server at your vault folder when it asks (or in its config). Run claude mcp list to confirm both are connected.
obsidian-mcp lets Claude read and write notes across the whole vault; qmd is the precision tool for pulling notes by path or tag. Together they cover full-text and structured retrieval.
For semantic "find related notes even when I don't know the exact title" search, install the Smart Connections plugin inside Obsidian itself (Community Plugins → Smart Connections). It builds the embeddings locally and complements the MCPs above.
Make your skills read it · the bridge
The MCPs let you query the vault in a chat. But the SkillTree skills you install don't read your Obsidian vault · they read a local knowledge/ folder (company.md, voice.md, offer.md, stack.md). Skills run fast against plain files, not a live vault.
The Brain Sync skill is the bridge. It pulls your Obsidian Brain down into that knowledge/ folder · mapping each note to the right file, showing a diff, and writing only what you approve. Edit in Obsidian, run sync, and every skill reads the current truth.
Change a fact in the vault → run "sync my brain from Obsidian" → your skills pick it up. That's the loop that makes "install a skill and it already knows my business" literally true.
The auto-memory layer
Alongside the vault, Claude persists what it learns across sessions in its memory directory:
~/.claude/projects/<project-hash>/memory/
├── MEMORY.md # routing document · keep under 200 lines
├── debugging.md # solutions to recurring problems
├── patterns.md # confirmed conventions
├── architecture.md # key architectural decisions
└── preferences.md # your workflow preferencesKeep MEMORY.md under 200 lines. It's a routing document, not a dump · detail goes in the topic files it links to.
Prefer a hand-rolled memory directory
If you'd rather own the structure instead of relying on Claude's auto-memory, create a memory/ folder in your project root with five files Claude reads and writes every session. Same principle, simpler shape:
memory/
├── decisions.md # every significant decision + reasoning
├── clients.md # deep client context
├── projects.md # active projects with current state
├── patterns.md # what's working and what isn't
└── open-loops.md # unresolved questions and pending itemsAt the end of each session, ask Claude to update the relevant files. Over a few weeks this becomes the most accurate picture of your business anywhere.
The session rhythm
- Orient · Claude reads CLAUDE.md and queries the relevant vault notes
- Work · full session with context loaded
- Persist · Claude writes what it learned into the right memory file
The persist step is where almost everyone drops off. It's also where all the value compounds.
End every session with: "Update my memory files with anything you learned today · new decisions, patterns, client context." Ten seconds of typing, and the next session starts smarter. Skip it and your vault stops improving the day you stop maintaining it by hand.
Common mistakes
- Treating CLAUDE.md like a settings file. It's a teaching document · write it like you're onboarding someone who remembers everything forever.
- Dumping everything into one big file. Atomic notes with prose titles are retrievable. Monolithic files aren't.
- Skipping the persist step. Without it, your session memory never improves.
- Not feeding Layer 3. The vault only knows what you put in it. Start with your last three most useful recordings.