Build a personal “second brain” where Obsidian is the storage (plain-text notes that link into a graph, on your own machine) and Claude is the brain on top — reading the whole vault, filing new information, linking it, and answering across all of it. Because it is just text files, it is not locked to any one model: point a different LLM at it next year and it still works. This is the LLM Wiki pattern Andrej Karpathy popularized in April 2026 — unlike a chat history that rots, the vault gets sharper every day. This vault, Brain, is a built-out version of exactly this.
🛠️ Prerequisites
Claude Desktop with a paid plan (Pro $20/mo) — the free tier won’t run Claude Code.
Obsidian (free) with a vault created on your machine.
The Local REST API community plugin in Obsidian (provides the API key Claude connects through). Obsidian must stay open while connected.
đź“ť Instructions
Install Claude Desktop, sign in (paid plan), open the Code tab — that’s Claude Code, the version that reads/writes local files.
Create the Obsidian vault — Create new vault → name it (e.g. “brain”) → pick a folder. That folder is your second brain. Learn the core mechanic: typing [[goals]] makes a link; notes pointing at notes form the graph.
Open the door — Obsidian Settings → Community plugins → Browse → Local REST API → Install → Enable. Copy the API key from its settings (the long string only — drop the leading Bearer).
Test with: “list every file in my Obsidian vault.”
Load yourself into the brain — don’t type your profile by hand; make Claude interview you one question at a time (who you are, this year’s goals, how to talk to you, strengths/weaknesses, current projects), then write it all to a CLAUDE.md at the vault root so it loads every session. (This vault’s root CLAUDE.md plays the same role.)
Create a project with structure — one folder per area, each with Inputs / Process / Outputs / Feedback subfolders and its own project-level CLAUDE.md (what the project is, its one goal, Claude’s role). Ideas land in Inputs, Claude works in Process, finished work goes to Outputs, metrics to Feedback.
Open a single project as its own vault (Manage vaults → Open folder as a vault → Trust) so Claude sees only that job’s CLAUDE.md. The big vault plans; a single project ships.
Build skills — anything done more than once becomes a saved markdown skill in the project’s skills/ folder, with a clear name + when-to-trigger description. Then “run the [name] skill” executes it your way. (Same idea as this repo’s Agent Skills and .claude/skills/.)
Wire in live data — connect read-only MCP servers, e.g. Google Calendar:
claude mcp add google-workspace uvx workspace-mcp --tools calendar
Add Gmail/Slack/Notion the same way; grant read-only wherever possible.
Put it on autopilot — schedule a daily task (Claude Desktop Schedule → New task) that files new Inputs, links them, flags stale notes, and writes a 3-line “what changed overnight” summary.
Outcome
Claude stops forgetting you at the end of each session. You get a vault holding everything, a model that resumes exactly where it left off, and a graph that grows itself — all plain text, all yours, portable across models. You stop re-explaining yourself and stop rebuilding context from memory every time.
One rule you never break: keys, not prompts. Telling an agent "don't delete this" is a suggestion, not a safety control. If it can delete a file or send an email, assume one day it will. Enforce access at the permission level — read-only, scoped keys — never with words in a prompt.
đź“– Further reading
LLM Knowledge Bases — the Karpathy LLM Wiki methodology this guide operationalizes.
OpenKB — a CLI that automates this whole compile-to-wiki loop; Second Brain Design — the PKM design behind this vault; Brain — this digital garden.