Your notes will outlive the app they're in, or they won't

Steph Ango, the CEO of Obsidian, wrote a short essay last year called "File over app." The argument is simple: apps are ephemeral, files can last, so build your digital life around formats you control instead of tools you rent access to. He put a number on it. If you want your writing readable on a computer from the 2060s or 2160s, it needs to be readable on a computer from the 1960s. Plain text clears that bar. A proprietary database behind a login does not.

I read that essay in August, a few weeks after reading his companion piece on how he actually organizes his own vault: almost no folders, heavy internal linking, notes that live in the root unless they're clearly a reference or a clipping from someone else. The structure isn't imposed from above. It emerges from thousands of small linking decisions made in the moment, the same way a city grows from thousands of individual choices about where to put a door.

Neither essay mentions AI. Both of them explain, better than most of the AI-and-knowledge-management writing I've read this year, why the format war matters now more than it did five years ago.

Here's the mechanism. An AI agent that's supposed to help you think, write, or recall something is only as good as what it can actually get to. If your archive lives in a format only one company's reader can parse, the agent is stuck asking that company's API for permission, at that company's rate limit, in that company's schema, for as long as that company exists and keeps the lights on. If your archive is plain markdown in a folder, any agent, today's or one built in 2031, can open it directly. The durability argument was always about outliving the app. It turns out it's also about which tools get to use your knowledge in the meantime.

Read-it-later apps and bookmark managers are a specific, unglamorous test case for this. They're where most people accumulate years of judgment: the article that changed how you think about something, the thread you meant to act on, the recipe you'll actually cook someday. That archive is valuable precisely because someone spent time deciding it was worth keeping. And it's usually the most locked-up data most people own. Full text sits behind a proprietary reader. Tags live in a schema only the app understands. Export, when it exists, produces a JSON dump that technically contains your data and practically requires a script to make sense of.

I've watched this pattern show up in my own save library while building Saive. A save isn't just a URL. It's the article text, my highlights, whatever note I attached, the tags, the date I decided it mattered. If any one of those pieces only exists inside the app's own format, the save is worth less the moment I want an agent, or a future version of myself, to actually use it for something. That's not a hypothetical. It's the difference between an agent that can summarize my last month of reading in one query and an agent that can't see past the app's front door.

The industry's answer to "AI plus your notes" has mostly been about wiring things up: plug your vault into a model, add a chat interface, call it AI-native. That's the easy half. The harder, less marketable half is making sure what you're wiring up is actually worth reading once it's connected. An agent pointed at ten thousand items you saved and never revisited, stored in a format it has to fight to parse, doesn't produce insight. It produces a longer, faster-generated version of the same clutter.

A few bookmarking tools have started marketing "keep it in markdown" as a headline feature this year, which is a good sign. It means the argument is landing outside the Obsidian crowd it started in. But storing things as markdown solves the format problem, not the judgment problem. Ango's vault essay is really about the second one: he uses folders sparingly, links aggressively, and revisits old notes on purpose, at random, specifically so the structure keeps reflecting what he actually believes rather than what he believed when he first saved something. A bookmarking tool that stores clean markdown but never asks you to revisit anything has fixed the easy problem and left the hard one exactly where it was.

That's the test I'd apply to any tool claiming to be AI-ready for your knowledge, bookmarking app or otherwise. Export everything tomorrow. Do you have a folder of files you or any agent could actually read and reason over, or do you have a pile of IDs that only make sense inside the product that generated them. Plain text was always the safer bet for the writer. It's becoming the safer bet for the AI reading over their shoulder too.