What to look for in a “second brain” — and why most rent you back your own mind.

Every comparison chart for these tools leads with speed, price, or how good the summaries are. Before any of that, there is one question that decides whether you still have your own notes in five years — and almost nobody leads with it.

A second brain, parked in someone else’s account

A second brain is supposed to be the place you put years of thinking so you can stop carrying all of it in your head. That is exactly why where it physically lives matters more than anything else on the comparison chart. If the notes live inside someone else’s account, you have not built a second brain. You have leased shelf space in someone else’s building, and the rent is a subscription to your own memory.

That is not an accident. Being the place your thinking lives is the business model — once years of notes are inside, leaving costs you everything you built there. So before speed, before price, before how good the summaries sound, there is an earlier question: whose building is this, and what happens the day you want to move out?

Six questions to ask any tool

Ask these before the demo, not after a year of notes has already moved in.

1. Can I open the files without the app? If reading your own notes needs that company’s software installed and running, the notes are hostage to the software’s survival, not just to your subscription.
2. Can I take them to a different model tomorrow? Model quality resets every few months, and notes that only speak one model’s private format keep you from riding that curve.
3. Does the tool tell me what each answer cost and where it ran? A tool that never shows its own meter is asking for trust it has not earned.
4. What leaves my machine, exactly, and when? A vague answer here — “we take privacy seriously” — usually means everything leaves, by default, on every question.
5. What happens to my notes if the company or the plan disappears? A second brain is a decade-long bet, and most of the companies building one have not been around a decade yet.
6. Does it get smarter from my corrections, or only from more of my data? Those are two different products. One compounds your judgment. The other just compounds its training set.

How the usual answers line up

Line the categories up against those six questions and the same pattern shows up: not dishonesty, just the default shape of a hosted product.

Hosted AI notebooks — NotebookLM is the best-known example — ground every answer in the documents you upload and cite the exact passage it came from, which is genuinely useful. But the documents live in the vendor’s account, read by whichever model the vendor runs that quarter, and the notebook does not travel to a different model or a different company if you leave.

OS-level and chat-app memory — ChatGPT’s memory feature is the most visible version — is convenient because it is ambient: it remembers without you filing anything. That convenience is also the limitation. The memory belongs to that product’s account, in that product’s format, and “export” usually means a data dump you could parse yourself, not a vault built to be opened elsewhere.

Always-on recorders — wearables and desktop tools that capture audio and screen activity continuously — solve a different problem: you never decide in the moment what is worth keeping, because everything is kept. That is also the weak point. Months of continuous capture is a lot of raw material, most of it low value, and storing that much of it is heavy enough that it typically lives in the vendor’s cloud, not on your device.

Note apps with AI layered on start from something you already trusted: your own writing, in an app you used before AI arrived. The AI layer usually runs as a round trip through the vendor’s servers, so asking a question of your own notes sends those notes somewhere else to answer it.

Where chanio lands on the six

Run the same six questions against chanio, honestly.

Files on disk: yes. Plain Markdown in a folder you chose, readable in any text editor — no app running, no account to sign into.

Model-portable: yes. The files were never written in a format that ties them to one model. Point a different one at the same vault next year and keep going.

Shows what each answer cost and where it ran: yes — that is close to the point. Three tiers do the work: free on-device answers now, harder questions held for tonight, one nightly frontier batch — and the meter shows which tier answered and what it cost.

What leaves the machine, exactly: only the questions the local pass held as too hard, and only the passages you approve, sent once, batched into that single nightly run. Not the vault. Not continuously. Not by default.

If chanio disappears: you keep the Markdown. It was never inside chanio to begin with — it was already on your disk, and it stays there.

Does it get smarter from your corrections: yes, though coarsely at first. Correcting the local pass changes how it sorts and what it holds for the nightly batch, and that gets sharper with use — it does not arrive sharp on day one, because a small model doing first-pass triage is not a frontier model.

The honest part

Hosted tools are more polished, and they start faster. Opening a browser tab beats installing something and pointing it at your own files, and the gap in finish is real, not marketing. The extra step on chanio’s side is the price of the files never leaving, not a hidden virtue.

A small model reading on your own machine is weaker than a frontier model doing that same reading. That is not a footnote to bury — it is why the loop has three tiers instead of one, and why the genuinely hard question waits for the nightly batch instead of a confident, weaker answer on the spot.

Owning the files means owning the backup. A vendor’s cloud backs itself up quietly, bundled into what you already pay for. A folder on your disk does not, until you set that up yourself — a small chore the hosted version never asked of you.

What to do this week

Pick one export — whichever tool is currently holding the largest share of your last year of thinking — and run it through the rescue tool, which turns a ChatGPT or Claude export into a wikilinked Markdown vault, parsed entirely in your browser, with nothing uploaded. Let chanio run one real night against it: today’s hard questions held and batched, a brief waiting when you wake up. Read that brief before you read another tool’s feature list. Then decide with the six questions in front of you instead of a demo.