# Don't put AI in your notes app. Connect your notes to your AI. Imagine opening a note two years from now. It's about the kitchen renovation, or a client, or a book you loved. Halfway down there's a paragraph that sounds a bit smoother than the rest. Did you write that? Or did the app add it one afternoon while being helpful? That worry is why a lot of people keep AI away from their notes entirely. And they're right about the danger. They're just aiming at the wrong thing.

Quick answer

Most of the writing you didn't ask for comes from note apps with AI built into them, quietly summarizing and tagging in the background. There's another way round. Keep the notes app simple, with no AI of its own, and connect the assistant you already talk to, like Claude or ChatGPT. That's what Hjarni is. It doesn't run an AI or send your notes to one on its own, so writing arrives because you asked for it in a conversation, and every change is saved with the name of whoever made it so you can undo it.

## Slop comes from software that writes on its own Think about where the mess usually comes from. You can certainly make some yourself, by asking for summaries you never reread. But the bulk of it tends to come from features that run whether you asked or not. Summarize every note. Suggest tags for everything. Find connections across the whole collection, overnight, forever. Each one sounds useful in a menu. Together they produce writing at a rate no person can review. And because it keeps arriving whether you read it or not, it piles up. A year later your search results are full of text nobody wrote on purpose. So the choice people think they're making, AI or no AI, isn't the real one. The real one is whether something is generating text when you're not there. ## There are two places the AI can sit It can sit inside the notes app. That's the usual approach: a note app with its own AI built in, buttons that write things for you, and something quietly tidying in the background. Or it can sit where you already are. You're talking to ChatGPT or Claude anyway. That conversation is where the thinking happens, so your notes don't need a second AI of their own. They just need to be somewhere that assistant can look things up and write things down. That's a one-time bit of setup: you tell ChatGPT or Claude where your notes live, roughly the way you'd let an app see your calendar. After that it can read them and add to them while you chat. [Here's what that connection is](/docs/what-is-mcp), if you want the longer version. That second arrangement is what we built. Being precise about it, because it's unusual: **Hjarni doesn't run an AI, and never sends your notes to one on its own.** No summarize button. Nothing reading your notes overnight. There's no AI service anywhere in the app to send them to. It does do a little bookkeeping by itself, and it's the boring kind. Rename a note and it tidies the links pointing at it, so they don't go stale. That's a find-and-replace, not a sentence anyone wrote. Otherwise your notes just sit there being notes. The thinking happens on the other end of a conversation you started. ## Which changes what shows up, and when Here's what that means day to day. Writing lands in your notes because you were in a conversation and asked for it. You research a supplier and ask the assistant to save what it found. You finish a call and ask it to write up what was decided. You correct something it got wrong and ask it to fix the note. You're in every one of those. Not because we built a rule that forces it, but because that's the shape of the thing: there's no feature sitting in the app with its own reason to write. Two honest exceptions. You start with a handful of starter notes, so the place isn't empty on day one, and you can delete them. And an assistant you've connected can write whenever it's running, not only while you're watching it. Which is the argument for the next part. ## And when it does write, it signs its work The other half is being able to check. Every note in Hjarni [keeps its own history](/changelog/note-history-and-provenance). Every version is saved with the name of whoever made it, edits from Claude or ChatGPT get a badge so they stand out from yours, and you can see exactly what changed and put it back with one click. Nobody has to remember to label anything. People do try to keep track by hand, writing "AI wrote this" at the top of a note or keeping generated notes in a folder of their own, and it works when you remember every single time. This just does it for you. To be straight about the limits: it won't stop an assistant writing something wrong, and it won't read the changes for you. What it takes away is the part that made people nervous, that a mistake becomes permanent because you can't find it again. ## What assistants actually do when you connect them We looked at [a month of real usage](/blog/what-ai-does-with-write-access): 108,613 things assistants did with people's notes, from looking something up to writing something down to routine checking around the edges. They changed notes that already existed 3.7 times more often than they made new ones. So the nightmare version, an assistant flooding you with hundreds of new files, isn't what usually happens. Being fair about that number: it counts how often they edited rather than created. It doesn't tell you whether any given change was a good one, or what was in it. That part is a judgment call, and it's why the history is there. ## Telling it what you want, and what that's worth You can set rules for how an assistant writes: what to include, how to format it, which folders to stay out of. [You write them in plain English](/docs/ai-instructions), and assistants read them constantly. In that same month, checking the rules was the fourth most common thing they did. But be clear about what a rule is. It's a house rule a well-behaved app follows, not a lock. A connected assistant can reach everything your account can reach, and it [can't currently be limited to a single folder](/docs/privacy-and-permissions). So the promise isn't that it can never write somewhere you'd rather it didn't. It's that if it does, the change has a name on it and you can take it back. If that side of things is what's stopping you, we wrote out [exactly what a connected assistant can and can't do](/blog/is-it-safe-to-give-chatgpt-access-to-your-notes). ## Let it do the boring parts The split we like is the one we wrote about [when Andrej Karpathy described keeping a wiki an AI looks after for you](/blog/karpathys-llm-wiki-is-right): let the assistant do the filing, you do the thinking. In practice that's small stuff. Writing down the decision you just made. Fixing a detail that went out of date. Adding this week's entry to a list you keep. The essays, the opinions, the connections between ideas: keep those yours, and say so in your rules. That's where the value of your notes comes from in the first place. Reading is the easy half, and it's most of the reason to connect anything. The one call worth making up front is what to connect, since anything an assistant reads becomes part of your conversation with that company. ## If you'd rather have no AI writing at all You can get most of the way there. Tell assistants to read and never write, then check the history now and then to see whether they listened. What you can't do today is make it impossible, which is the honest version of that answer. But if what you want is files on your own computer, no online account, and nothing generated anywhere near them, that setup is genuinely good and we're not going to pretend otherwise. The honest comparison is in [Hjarni vs Obsidian](/compare/hjarni-vs-obsidian). For everyone else, the question was never whether to allow AI. It's whether anything is writing while you're not looking. Keep that out of your notes app, and let the assistant you're already talking to do the rest. [Give your AI a memory](/registration/new), keep the byline. Write once, you both remember.