All video notes
Video notes · Notes and PKM

Build an Obsidian BACKEND Not a Second Brain!

Video by Eric Michaud on YouTube · 8 min · uploaded October 2026

Notes and analysis by BrainTube, written from the video. Not affiliated with Eric Michaud. Watch on YouTube

In short

The video argues for treating an Obsidian vault as a plain-file backend that agents read and write, rather than a notes pile you never revisit. Its best ideas are a hard split between human and machine folders, an agents file in every folder, and metrics written by scripts instead of the model. The rules are enforced only by instructions, one rule contradicts another, and the dashboard shown is a paid add-on.

The argument

Michaud's point is that most vault setups only collect, so the information goes stale. He wants the vault to feed him instead. Because it is just markdown in a folder, any agent can read and write it, whether Claude Code, Codex or whatever ships next, and nothing has to be exported. The design has four parts. The vault is split into a human side the agent may not edit and a machine side where it can do anything. Every folder has an agents file that explains what lives there and what may be touched, so the agent reads a map instead of the whole vault. Scheduled collector scripts write each platform's numbers into the front matter of the daily note, so the data is consistent and never passes through a model. A findings file records what worked last time, and the agent must read it before the next run. With daily notes as columns and properties as rows, he says, the vault becomes a database you can build on.

Key ideas, with timestamps

  1. 1:01

    Plain files outlive any one agent

    He addresses critics who say a vault is just a folder of text files and turns it into the main argument. Claude Code, Codex, other agent harnesses and whatever comes next can all read and write the same files. Nothing needs exporting and nothing sits behind a subscription.

  2. 1:27

    Separate human and machine

    He keeps his own notes apart from the agents' output because models tend to summarise and merge things you wanted left alone, and you cannot be sure they kept your intent. The human side holds daily notes, tasks, projects, people and context; agents may move or copy there but not write or edit. The machine side holds scripts, logs, research and workflows, with full permissions.

  3. 2:13

    An agents file in every folder

    Each folder, and the vault root, has an agents markdown file that acts as the instruction manual for that section: what it is for, what can be touched and how to work there. The root file works as a map, so an agent can go straight to the right folder and read only that folder's rules.

  4. 2:36

    Agents log their own work

    The single most important rule he gives his agents is to log everything they do in today's daily note. He no longer journals what was done; the agent fills it in. The daily note becomes the activity record for both him and the agents.

  5. 2:52

    Scripts, not models, write the numbers

    Each platform has its own collector script. YouTube comes in through its official API, and others such as his community platform, Instagram and TikTok run on a schedule, writing values into the daily note's front matter. Because the data skips the model, it is collected the same way every day. At 3:38 he frames each daily note as a column and each property as a row.

  6. 4:17

    Do not dump the whole vault into context

    He calls loading the entire vault into every chat the common mistake, comparing it to handing someone a stack of papers and expecting them to remember the middle. The agents files let the model read the map, then only the workflow and notes the task needs.

  7. 5:06

    A findings file per workflow

    His animated reels folder holds the assets and instructions, plus notes from every past run: performance numbers, which topics landed, which jokes worked. That goes into a findings file the agent must read before the next run, so each attempt starts from what was learned last time.

  8. 6:26

    A dashboard that can act

    The front end he opens with reads the same files and runs the same scripts as his terminal setup. A section called needs you suggests the next high-leverage task from the logged history, and because Gmail is connected it can draft and send email from there. At 8:10 he says this app is available through a paid membership.

What it gets right

The human and machine split is the most useful idea here, and it is uncommon. Many agent vault setups let the model rewrite your notes, and the result is a vault you can no longer trust as your own record. Giving the agent its own folder, where mess is fine, protects the part that matters.

Using scripts for numbers is also correct. A follower count or view total should come from an API, not from a model reading a screenshot. Deterministic collection makes trends comparable day to day.

The agents file per folder and the findings file are both cheap and effective. One keeps context small and relevant, and the other gives a repeated workflow a memory of what worked. His warning about dumping the whole vault into context matches what is known about long contexts: material in the middle is used less reliably than material at the start or end.

What it leaves out

  • The rules contradict each other. Daily notes are listed on the human side, which agents may not edit, yet the main rule is that agents log everything into today's daily note. Presumably agents append to a set section, but the video never says how that boundary works.
  • Instructions are not permissions. The human side is protected only because an agents file says so. A model that ignores or misreads it can still edit those files. Real enforcement needs file permissions, a tool allowlist or hooks that block writes to those paths, and version control as a backstop.
  • Findings files rot. A findings file that only grows will collect lessons that were wrong or no longer apply. It needs pruning and dates on entries, or the agent will keep learning from old runs.
  • Front matter as a database has limits. It works for a handful of daily numbers. For many metrics over years, querying hundreds of files is slow and fragile compared with a small real database, and query plugins such as Dataview only go so far.
  • Sending email from a dashboard is an action with consequences. The video does not say whether drafts are reviewed before they go out.
  • No build steps. It is an overview. The scripts, agents files and dashboard are not shown in detail, and the app is behind a membership.

How this applies to memory across AI tools

The argument at 1:01, that plain files can be read by any agent, is the right foundation for memory that is not tied to one vendor. It still assumes each agent runs on the machine with file access. Chat apps such as ChatGPT or Claude on the web cannot read a desktop folder, so the vault's memory stops at the terminal. See AI memory portability for why the format matters, and persistent AI memory across tools for reaching more than one client.

His split also suggests where outside material belongs. Your own notes stay on the human side; sources you watch and read are reference data. BrainTube handles that layer: 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 alongside a local vault.

Who should watch it

Watch it if you already run agents against an Obsidian vault and want design rules rather than a tutorial. The whole method is in 1:27 to 5:36: the human and machine split, the agents files, logging to the daily note, script-collected metrics and the findings file. The dashboard from 5:53 is a demo of his own app and can be skipped, as can the membership pitch at 8:04. Skip the video if you want setup steps or code; it describes the structure but does not build any of it.

Frequently asked

Can Obsidian be used as a database for AI agents?
For small, regular data it works well. Agents can read and write markdown files, and front matter properties give each note structured fields that scripts can fill in. For large or long-running datasets, a real database is faster and more reliable to query.
How do I stop an AI agent from editing my own notes?
Keep your notes in a separate folder from the agent's output and say so in an instructions file such as AGENTS.md. Instructions alone are not enforcement, so also use your agent's permission settings or hooks to block writes to that folder. Keep the vault in git so any unwanted edit can be reverted.
What is an AGENTS.md file in a notes vault?
It is a markdown file of instructions that coding agents read when they work in a folder. In a vault, one at the root can describe the overall layout, and one in each folder can say what lives there and what the agent may change. This lets the agent load only the context a task needs.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

Go deeper

More video notes

Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.