I keep saying “my vault lives in Obsidian.” It’s a convenient lie.
Nothing of mine actually lives in Obsidian. My notes live in a folder on my disk. Obsidian is one of several apps that happen to open that folder. The day I understood that distinction properly, the whole lock-in question — the one I wrote about in How I built an Obsidian vault no single AI tool can hold hostage — snapped into focus.
What we casually call “Obsidian” is three physically separate things — plus a fourth layer Obsidian never claimed. And that unclaimed layer turns out to be the most important one.
The four layers, unbundled
Storage. A folder on my disk. That’s the layer that remains when everything else is switched off — no servers, no account, no app. If Obsidian the company vanished tonight, the folder would not notice.
Format. Markdown, YAML frontmatter, wikilinks. An open standard any language can parse. A Python script, a shell one-liner, an AI agent — they all read my notes without asking anyone’s permission.
UI. Obsidian today. But VS Code, iA Writer, or Zed open the same folder tomorrow. The interface is a preference, not a commitment.
Agent. This is the layer that was never part of “Obsidian” at all — and that’s the point. Claude Code, then Codex, then back to Claude Code: I swapped my agent twice this year, and as I described in the vault-hostage article, neither swap touched how I work. An agent can slot into my system only because the first three layers are open. Obsidian didn’t build this layer. It never got in its way.
The genius here isn’t design — it’s the absence of design. Nobody fused these layers together, so each one can be replaced without touching the other three. Leaving Obsidian tomorrow wouldn’t move a single note. Changing agents doesn’t change the folder. Migrating from Mac to Linux means dragging a directory.
Integrated platforms take the opposite bet
Notion is the example I know best — I’ve written about why it isn’t my go-to tool — but this isn’t really about Notion. It’s about a whole category: the integrated, all-in-one platforms. Evernote played this role a decade ago, Notion plays it today, and something else will play it tomorrow. The pitch never changes: one product, everything inside.
Through the four-layer lens, “everything inside” means the layers are fused. Storage is their database. Format is their proprietary data model. UI is their app. Agent is their built-in AI. You cannot replace one without replacing all of them — and when you try to leave, the export drops part of your structure on the floor. In Notion’s case: databases flatten, links break, properties evaporate.
Let me be clear: fusion has real benefits. The UX is polished, real-time collaboration works out of the box, and nothing needs configuring. For a team, trading sovereignty for convenience can be a perfectly rational deal.
But it’s a deal. And most people sign it without reading the terms.
The new comparison axis: can an agent get in?
For years, note-app comparisons ran on the same criteria — editor quality, databases, collaboration, price. In 2026, a different question decides everything: can an external agent read, write, and operate on your system?
For my vault, the answer takes one sentence: any process on my machine can open the file directly. No API. No auth token. No rate limit. The agent isn’t a client calling a service — it’s a process reading the same files I do.
For a fused platform, the answer runs through the vendor: an API token, rate limits, a proprietary data model to convert — when an API exists at all, and only for what the vendor chooses to expose. The built-in AI can do plenty, but it’s their agent, on their infrastructure, with their models and retention policy. It cannot run offline. It cannot run on a local model. It cannot be a script I wrote on a Sunday morning.
That’s the real dividing line — not capability, but ownership of the agent layer. Agents improve every few months, and each generation slots straight into an open stack the day it ships. On a fused platform, you get better agents exactly when, and only when, the vendor decides you do.
The honest trade-offs
Four separate layers means four separate decisions, and nobody holds your hand across them. Sync is my problem. Mobile is rougher. There’s no real-time collaboration worth mentioning. And migration is never free even with plain text — moving a vault still costs broken links and cleanup hours. The four-layer design lowers the exit price; it doesn’t abolish it.
If you want software that works the moment you open it, and you never plan to point a script or an agent at your notes, an integrated platform will serve you well. And that’s okay.
Conclusion
I no longer evaluate note-taking tools by their feature lists. I count layers.
The robustness of a knowledge system is the number of layers you can replace without losing anything — four in my vault, none in a fused platform without paying the export tax. That number is not a technical detail. It’s how much of your system you actually own.
Count yours. How many layers could you swap tomorrow without losing a note?