Building an AI Brain From a Folder of Markdown Files

Every AI conversation starts from zero. You explain your project, re-share context, and hope the model remembers what you told it last week. I wanted something different — a persistent workspace where Claude already knows what I’m working on, what’s due today, and how I like to operate.

So I built one out of Markdown files.

The Concept

A folder of Markdown files that acts as a shared brain between me and Claude. I call mine LCARS. It holds my projects, daily plans, decisions, and preferences — all in plain text. When I open a Claude Code session and point it at that folder, Claude reads everything and picks up where we left off.

The key insight: Claude’s 1M token context window is big enough to just load everything. No retrieval system, no embeddings, no clever indexing. Just Markdown and a flat read.

The Stack

Three layers: the vault, the sync, and the agents.

The Vault — a directory of Markdown files organized by project:

• AGENTS.md — tells Claude how to behave (with CLAUDE.md as a one-line @AGENTS.md pointer so every agent picks up the same rules)
• USER.md — my preferences and context
• TODAY.md — the day’s prioritized tasks
• projects/ — split into work/ and personal/, each project with a brief.md as its canonical doc

The Sync — Obsidian Sync at $5/month. Keeps the vault synchronized across every device. I use Obsidian to view and manually edit files. Check a task off on my phone, it’s checked off everywhere.

The Agents — Claude Code sessions pointed at the vault folder. This is the daily driver: Claude reads the full vault on session start — every brief, every todo, every decision log — and operates with full context from that point forward. Planning, writing, project management all run through it.

Multi-Machine Setup

Because the vault syncs across machines, I can run Claude Code on more than one computer at once. Both read and write to the same files, changes propagate through Obsidian Sync.

• Primary machine — Claude Code does the focused work: planning, writing, project management.
• Mac Mini (secondary) — when a task needs the web, I hand it to Claude Cowork here for browser automation — the Chrome plugin or computer use, doing research or web workflows so they don’t tie up the main session.

Two machines, one brain, zero conflicts.

Mobile Access

On my phone I have Obsidian installed with the same vault synced — full read/write access to every file. I can review briefs, check off tasks, or scan my daily plan from anywhere.

I also use Claude Dispatch on my phone, pointed at the Mac Mini. Since the Mini has full vault access, I can chat from my phone and have Claude make quick edits, look things up, or kick off a task as if I were at that machine. And when I need the full project context from my phone, I’ll start a remote Claude Code session with /remote-control and pick up right where the desktop left off.

The Daily Loop

TODAY.md is the operational heartbeat:

1. Claude generates a prioritized task list each morning, pulled from every active project
2. Throughout the day, I check things off — in Obsidian or Claude does it after completing work
3. Claude syncs completions back to project files and generates the next day’s plan

One document that always answers “what should I be doing right now?” — and Claude always knows the answer too.

Rough Edges

A couple of trade-offs come with this setup:

• Obsidian must be running on the secondary machine at all times for sync to work. There’s no headless setup — just the app, always open.
• Dispatch is context-light. When chatting from my phone, Dispatch can reach the vault and make changes, but it doesn’t load the project context a full Claude Code session does. I have to be more explicit — “go read the Avolve brief and update the timeline” instead of just continuing a conversation — or fall back to a /remote-control session when I need the full picture.

Why It Works

Three assumptions the whole thing rests on:

• Context windows are big enough to load everything. The vault fits comfortably in 1M tokens.
• Markdown is the right format for human-AI collaboration. Readable by both humans and models with zero translation.
• Sync solves the multi-device problem. Obsidian Sync at $5/month is rock solid.

What I like most is the leverage. Every file I add makes every future conversation smarter. A decision I documented three weeks ago saves me from re-explaining it today. A project brief I wrote once becomes permanent context that Claude carries forward. The longer I run it, the less I have to repeat myself.