AI second brain: notes your agents write and you can trust

An AI second brain is a personal knowledge base where AI does part of the work: it captures what you read and hear, writes and links notes, and finds the right note when you need it. You still decide what's worth keeping and what's true.

The term builds on Tiago Forte's Building a Second Brain, a method he also published as a book in 2022. Forte describes a second brain as “an external, centralized, digital repository for the things you learn and the resources from which they come”. His method has four steps, which he calls CODE: Capture what resonates, Organize it by how actionable it is, Distill it into short summaries, and Express it as something you make. He organises notes with PARA: Projects, Areas, Resources and Archives.

Every one of those steps used to be manual, and keeping up with them is where many second brains stall. AI changes the economics. Here's what it takes over, and what can go wrong at each step:

Most tools sold as an AI second brain fall into one of two families. They solve different problems, and the right one depends on how you work.

Capture-and-chat apps

You throw things in (text, voice, links) and the app organises them. You get answers by chatting with your notes. The apps in this family come in a few flavours:

• Capture apps with an agent. Take voice notes, web clips and text, file them automatically, and let you ask an assistant to draft, summarise and plan from what you've saved.
• All-in-one assistants. Bring notes, email, calendar and tasks together, with an assistant that searches, schedules and creates across them.
• Source notebooks. Answer questions from material you upload, such as PDFs, web pages and videos, with inline citations.
• Workspaces with built-in AI. Answer questions from your pages and connected apps, and let their agent create and edit pages.

Good at: fast capture, including on the go; no setup; asking questions in plain language. Weak at: your notes live in the vendor's cloud and format, the AI in the app is the only one that works on them, and you rarely see exactly what it changed.

Files-first setups

Your second brain is a folder of plain markdown files on your computer. Instead of an assistant built into a notes app, you point a general agent at the folder: Claude Code, Codex or Cursor. The agent reads sources you drop in, writes and updates notes, maintains links and an index, and answers questions by reading the files. Andrej Karpathy described this pattern in his LLM Wiki note (April 2026), and our Claude Code second brain guide walks through the setup.

Good at: you own the files; any agent or editor can work on them; every change is a diff you can inspect; you can use a stronger model the day one comes out. Weak at: capture on your phone is harder, you need to be comfortable running an agent, and nothing stops the agent from quietly rewriting a note unless you set that up.

A second brain is only worth having if you can rely on it. Once AI writes a large share of the notes, three things you used to know for free start to slip away:

• Which notes you wrote, and which the AI wrote. A note in your vault reads like your own thinking, even when an agent wrote it from a podcast transcript you never finished.
• Where a claim came from. A summary drops the source. A summary of summaries drops the nuance. Six months later you quote a number and can't find where it came from.
• Whether it's still true. Tax rules, prices, your own plans and opinions change. An old note sounds exactly as confident as a new one.

A concrete example: you ask an agent to add a long article to your notes on pensions. It writes a clean summary and, to keep things tidy, updates your existing note with the article's contribution limit. The article was from two years ago and the limit has since changed. Nothing looks wrong. The next time you plan, you plan with the old number.

The problem isn't that AI makes mistakes; people do too. It's that the mistakes arrive silently, in your own notes, in your own voice. Chat apps that cite their sources help at the moment you ask. They don't tell you which notes in your collection were never checked by you.

The fix is not to stop using AI. It's to make every note carry its own history, and to look at what the AI changes before you rely on it. Here's how to set that up with Knowtarium and the agent you already use. The notes stay plain markdown throughout, so every step also works with a text editor and git.

1. Keep it in plain files

Markdown in a folder you control. Any editor opens it, any agent can work on it, git or a sync service backs it up, and nothing is lost if an app shuts down. A PARA-style layout works well:

~/brain/
  AGENTS.md               # how agents write here (CLAUDE.md can point to it)
  index.md                # one line per note, kept up to date by the agent
  log.md                  # what changed, newest first
  sources/                # raw material: clippings, transcripts, PDFs as text
    2026-09-12-podcast-sleep.md
  projects/
    marathon-2027.md
  areas/
    health/sleep.md
    money/pension.md
  resources/
    books/four-thousand-weeks.md
  archive/

Keeping raw material in sources/, untouched, matters. It's what you check a summary against later.

2. Record who wrote each note

Put the provenance in the frontmatter, where both you and the agent can read it. The Open Knowledge Format (OKF), an open specification published by Google Cloud (v0.2 on GitHub), has fields for this: generated for who wrote it, verified for who checked it, sources for where claims came from and stale_after for when to look again.

---
type: Note
title: Sleep and training load
description: What I've learned about sleep before long runs.
tags: [health, running]
generated: { by: claude-code/2.1, at: 2026-09-12T20:04:00Z }
verified:
  - { by: human:alex, at: 2026-09-13T08:15:00Z }
stale_after: 2027-09-12T00:00:00Z
sources:
  - id: podcast-sleep
    resource: /sources/2026-09-12-podcast-sleep.md
---
 
The guest recommends a consistent wake-up time over extra hours in bed.[^podcast-sleep]
My own note: I sleep worse after runs that finish past 8pm.

Agents sign their work with their name and version (claude-code/2.1); you appear as human:alex. A note an agent wrote with no verified entry is one you haven't read yet, and you can see that at a glance.

In Knowtarium

Knowtarium shows this frontmatter as properties at the top of the note instead of raw YAML, and the graph shows what's gone stale. With review on, it shows who checked each note too, and the graph can colour notes by trust tier: unverified, machine-confirmed or human-reviewed. The tags field and any keys of your own stay exactly as you wrote them.

(opens full size in a new tab) (opens full size in a new tab)

3. Tell your agent the rules

Agents follow written conventions. Put them in AGENTS.md at the top of the vault (Claude Code reads CLAUDE.md, which can point to it). The AGENTS.md guide has a longer template.

# How to work in this vault
 
- Read index.md first. Update an existing note before creating a new one.
- Save raw material to sources/ unchanged, then write or update notes from it.
- Set generated.by to your name and version on every note you write or change.
- Never edit text under a heading called 'My own note'. Suggest changes instead.
- Cite a source for every claim that isn't my own opinion: [^source-id].
- Never add anyone to verified. I do that when I've read the note.
- Add a line to log.md for each change.

4. Make the agent cite sources

Every claim that isn't your own opinion gets a footnote keyed to a source id, like [^podcast-sleep]. When you doubt a sentence, you can open the source it points to. See how to make agents cite sources.

5. Review what the AI changes

Look at every agent change as a diff before you rely on it. With the vault in git, git diff shows exactly what changed. Pay most attention to edits of existing notes, not new ones: a changed number in an old note is easier to miss than a new page. When a note is right, add yourself to verified. How to review agent edits has a checklist.

In Knowtarium

Each change an agent saves in Knowtarium is signed and marked on the note, with a line under the title that links to the diff (See changes) and an Undo. If you'd rather approve first, turn on review for a folder: changes there wait as diffs, and approving one adds you to verified with the time, which agents can't do for you. In those folders, your own edits wait until you ask an agent to check them against the notes connected to them.

(opens full size in a new tab) (opens full size in a new tab)

6. Put an expiry date on facts that change

Give notes about prices, rules, tools and plans a stale_after date. Once a month, ask your agent for everything past its date, have it re-check those notes against their sources, and review what it proposes.

7. Keep your own thinking yours

Let the agent summarise, link and file. Keep your opinions and decisions in sections it may not rewrite. The point of a second brain is to help you think, and it can only do that if you can still find your own thinking in it.

Choose a capture-and-chat app if most of your notes start on your phone, you want to ask questions rather than browse, and you don't want to run an agent yourself. For many people that's enough.

Choose Knowtarium with your own agent if you already use an agent like Claude Code, you want your notes in an open format that outlives any one app, and the agent writes enough of your notes that you need to know which ones you've checked.

You can also mix them: capture in a mobile app, export to markdown regularly, and let an agent file it into your vault. To see what Knowtarium does, read the features and how it works with Claude Code, Codex and Cursor. For the shared, team version of the same idea, see company brain.

Knowtarium brings the files-first kind of AI second brain to an end-to-end encrypted web app: your notes are OKF markdown in workspaces and folders, your own agents work in them over MCP, and every note shows who wrote it (and, in folders with review on, who checked it). It's free to start, with 500 MB of storage and one workspace, and Starter ($5.99 a month) and Pro ($11.99 a month) add storage and workspaces.

Before you try it, here's what it doesn't do:

• It isn't a capture app. The web clipper saves a page's link and title (and your note) to your inbox, not the page's text, and there's no voice recorder. Your agent brings material in.
• It isn't a chat app. Knowtarium doesn't call an AI itself. The agents you already use (Claude Code, Codex, Cursor, Claude Desktop) connect over MCP with one command and get the conventions with their tools. Connectors for chat apps such as claude.ai and ChatGPT come later.
• There's no phone app yet. Knowtarium is a web app; a phone app comes later.