Claude just killed the note-taking app. Here is proof.

Data in files you control, model on top. The strongest idea is the separation between the folder (markdown instructions, documents, a SQLite file) and the AI working on it. If the structure lives in plain files, you can point a different agent at it later. That is a better test for any notes setup than features: can you leave without losing anything?

Inboxes with clear direction. One folder for what you hand to the AI, one for what it hands back, is a simple contract. It stops the agent from scattering outputs, and it gives you one place to review. Anyone can copy this in five minutes.

Check understanding before work. The plan-mode segment is the most honest part of the video. The agent guessed wrong about the team inbox, and the questions surfaced it. The 33:41 segment shows the same pattern: the agent made a reasonable but wrong filing decision, and a written rule fixed it.

Obsidian is not a lock-in example. At 53:23 he says Obsidian files are formatted so only Obsidian can use them properly. Obsidian notes are plain markdown files on disk. Wiki-links and some plugin syntax do not render elsewhere, but any editor, script or AI agent can read them, which is exactly what his own setup relies on. If anything, an Obsidian vault is an easy starting folder for his method.

Cost and scale. He mentions running the $200 per month Max plan and burning 21 million tokens during this session, and suggests the $20 Pro plan is enough to start (prices as of the March 2026 upload). OCR across 119 PDFs did not finish on camera. For a large archive, token cost and run time are real limits that a notes app does not have.

Reliability over months. The demo is one session. It does not show what happens when the CLAUDE.md file, agent definitions and database schema drift, or when two of the six to eight parallel Larrys he says he runs write to the same database. A notes app enforces a schema; an agent-maintained folder only follows rules you remember to write down.

Search is assumed, not shown. He says the database makes things easy to find, but never runs a real query against the scanned documents. Structured tables suit contacts and projects. Finding an idea inside a document is a different problem, usually solved with full-text or semantic search.

The setup is tied to one folder on one machine and one agent session. That works while you are inside Claude Code in that folder. It does not help when you switch to ChatGPT on your phone, or Cursor in another project, and want the same material available.

The portable version of his idea is to keep your saved sources in one place and expose them over MCP, so any client can query them. BrainTube takes that route for the reading and watching side: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. His local folder can still handle journaling and contacts.

For the trade-offs between local files and shared memory, see persistent AI memory across tools and what an MCP memory server is.