Claude forgets you. Your notes don't.

Stop renting your memory. Run a real one on your own machine in 5 minutes.

Claude is the smartest model on earth with the memory of a goldfish. This repo gives it a permanent one — built on the notes you already write in Obsidian, indexed by a local model, searched by meaning instead of keywords. No cloud. No subscription tier. No leash.

One file. Double-click. Done.

Quick install

# macOS
curl -L -O https://raw.githubusercontent.com/Coherence-Daddy/give-obsidian-a-memory/main/installer/install.command
chmod +x install.command && open install.command

# Windows
Invoke-WebRequest -Uri https://raw.githubusercontent.com/Coherence-Daddy/give-obsidian-a-memory/main/installer/install.bat -OutFile install.bat
.\install.bat

# Linux
curl -L -O https://raw.githubusercontent.com/Coherence-Daddy/give-obsidian-a-memory/main/installer/install.sh
chmod +x install.sh && ./install.sh

The installer brews the dependencies, pulls the local models (llama3.1, nomic-embed-text), copies the vault template, wires claude_desktop_config.json, and opens both apps. ~3–8 minutes total. See installer/README.md for the per-platform details and caveats.

What you get

• Obsidian — your notes editor, pre-configured with a clean coral-dark theme, color-coded folders, and a starter vault that knows where everything goes.
• Claude Desktop — wired to read your vault via MCP. Ask it anything in plain English.
• Ollama — your model runner. Two ways to use it:
  • ☁️ Ollama Cloud (recommended): free tier, frontier-quality models (gemma4:31b-cloud, glm-5.1:cloud, kimi-k2.6:cloud, qwen3-coder-next) running on Ollama's hardware. ~0 local disk. Sign up at ollama.com, grab a free API key, and the brain auto-uses it.
  • 🔒 Local (privacy fallback): llama3.1 + nomic-embed-text cached on your machine, ~5 GB. Air-gap-grade — nothing ever leaves the laptop. Pick this if total privacy beats answer quality.
  • The 🤖 Models button toggles between them on the fly.
• A vault template — 🏠 Home, Tools/, Projects/, _Claude_Ollama_Human/, and the conventions Claude reads to behave consistently.
• The full Local Brain system (opt-in during install) — Qdrant + BGE-M3 semantic search, eight Obsidian sidebar buttons (📋 Open Dashboard, 🔍 Ask Brain, 🔄 Sync Brain, 📥 Intake, 📂 Import, 🤖 Models, 📐 View Mermaid, 🎨 View Excalidraw), seven Claude Code slash commands (/ask, /sync, /intake, /import, /dashboard, /status, /diagram), and a FastMCP server so Claude Code can search the brain mid-conversation without you typing a thing.

The two flavors

	Lite	Full
Setup	Steps 1–4 of the installer	Steps 1–5 (say "yes" at the prompt)
Search	Claude Desktop reads files via filesystem MCP	Semantic search over BGE-M3 embeddings in Qdrant
Ergonomics	Type questions into Claude Desktop	One-click sidebar buttons + /ask slash command + auto-recall via MCP
Dependencies	Claude Desktop, Obsidian, Ollama	Adds Docker (for Qdrant + embedder) and a Python venv
Best for	"I just want Claude to know my notes"	"I want a real brain — searchable by meaning, queryable from anywhere"

Both share the same vault template; the Full flavor is a strict superset.

What this is not

• Not a SaaS. Nothing leaves your laptop.
• Not a memory feature. It's a brain — your notes, embedded once, queryable forever.
• Not a code repo. Source stays where it lives. The vault holds the map.

Walkthrough

The 16-slide visual setup guide is the canonical entry point:

→ coherencedaddy.com/tutorials/give-obsidian-a-memory

The same HTML is mirrored at presentation/index.html in this repo if you want to fork or run it offline.

Companion tutorial

If you want Claude Desktop + Claude Code paired with Ollama (no Obsidian, just terminal cost-cutting), see:

→ Coherence-Daddy/use-ollama-to-enhance-claude — Save 90% on Claude Code quota by routing through Ollama.

Repo layout

presentation/
  index.html              # the 16-slide tutorial (self-contained)
  cd-face-coral.png
installer/
  install.command         # macOS double-click installer
  install.bat             # Windows launcher
  install.ps1             # Windows worker (called by install.bat)
  install.sh              # Linux installer
  README.md               # per-platform notes + caveats
vault-template/
  🏠 Home.md
  Tools/                  # the 8 button .md files (Open Dashboard, Ask, Sync,
                          #   Intake, Import, Models, View Mermaid, View Excalidraw)
  Projects/               # empty; fills as you import
  _Claude_Ollama_Human/   # conventions Claude reads
  .obsidian/              # pre-configured plugins + coral-dark snippet
local-brain/               # the Full-flavor system (copied to ~/local-brain on install)
  *.command               # button launchers (ask, sync, intake, import, models,
                          #   view-mermaid, view-excalidraw, dashboard)
  src/brain_manager.py    # indexer + semantic search pipeline
  src/mcp_server.py       # FastMCP server exposing brain_search to Claude Code
  docker-compose.yml      # Qdrant + BGE-M3 embedder, bound to 127.0.0.1
  engine.conf, engine-prefix.sh, .env.example  # Anthropic ↔ Ollama toggle
  deploy-buttons.sh       # idempotent vault-deployment script
claude-commands/          # slash commands for Claude Code (~/.claude/commands/)
  ask.md, sync.md, intake.md, import.md, dashboard.md, status.md, diagram.md
walkthroughs/             # plain-English flowchart pages, one per button
  index.html              # 9-card grid; entry point
  ask.html, sync.html, intake.html, import.html, dashboard.html,
  models.html, diagrams.html, mcp.html
  _shared.css, logo.png

Why this exists

Every Claude user has hit the same wall: you teach it the context, it forgets, you teach it again. Anthropic's memory feature lives in their database, on their schedule, behind their pricing. This is the opposite — a local memory you control, that works the way you already think (writing notes in Markdown), with a model that finds them by meaning.

License

MIT — built by Coherence Daddy.