For fifteen years, vendors in the personal knowledge management industry have been selling you the wrong solution.
Evernote. Roam. Obsidian. Notion. Logseq. They’ve all been focused on the wrong problem: capture. Faster clipping, better linking, smarter tagging, backlinks, graph views that looked like constellations and said nothing about meaning. The solution was always “get it out of your head and into the system” because having it somewhere else was presumed to make it easier to find later.
The system held it, and that was the problem.
Anyone who has maintained a vault for more than two years understands this moment. Four thousand notes. There’s definitely a note that contains precisely what you’re looking for. You don’t remember where it is, which is why you can’t find it. Finding was supposedly solved ten years ago, or at least the finding was never the problem.
The problem is that synthesis never actually happened.
This is what a model can fill.
What actually changes
If you point a language model at your notes, you’re fundamentally changing the nature of the search. Linear search for filename matches, or even content matches, is replaced with natural language queries. You stop thinking about what terms and tags you attached to a note in 2023 and start typing what’s in your head, however incorrectly, and something relevant comes up.
Even better, it can see what’s common between them and will actually look at twenty of them at once, which is the point the original Zettelkasten promised to deliver and no human could manage because we could only hold three ideas in our heads at once.
Three ways to wire it
It depends on what you want. The paste-tier solutions are the worst, but also the easiest to get started with. You’re essentially copying your notes into a Project as context, and you’re done. It stops working well once your notes get to be more than a few hundred pages, at which point you’re probably maintaining a different mirrored copy anyway.
The connector tier solutions are probably what you want to start with. You install an integration, and you point Claude at your drive or Notion or wherever your vault is, and it can query it as needed.
The filesystem tier solutions use either Claude Code or Coworker (or whatever comes next) to mount a directory as a virtual vault. Now it can read and write, follow links, grep, everything. This is where you want to be if you have serious content. The only caveat is that you want to have git history on those files if you can, which is probably why you’re using Obsidian in the first place.
Where it breaks
It invents connections where there were none. Notes with similar language aren’t necessarily about similar things, especially if you’re chunking your notes. The model will happily synthesise a discussion of ideas it’s seen in the past, most of which were wholly unrelated to the thing you’re actually trying to think about.
The value of long note content decreases dramatically. The four-thousand-word note has to be chunked somehow, and the part that interests you in particular may not have been the one that was selected.
It has no awareness of time. Your 2019 opinions are right there in the mix with everything else, undated and unflagged. It may quote you verbatim from a note you wrote three years ago that you’ve since outgrown.
Writing back is where people get hurt. If you’re an agent with write permissions and you ask it to “edit this note”, it will. There shouldn’t be any surprises, but there can be. An incorrect edit of a note you rely on can be worse than no edit at all, especially if you don’t know it’s there.
The part nobody wants to sit with
The effort is not a bug. It is the entire point.
The reason you had to write the note in the first place was that it was hard to remember without it, which is a big reason retrieval practice is an established learning technique for all ages. Every system that made it easy to forget the note made it less valuable. Nobody complains about the effort it takes to remember phone numbers, but the counterargument is completely valid and mostly beside the point. The second brain was never supposed to be a phone book, and it’s always been a thinking tool. Synthesis is thinking, and if your “vault” only exists as a latent vector space in Virginia, then it’s training data for someone else’s first brain.
That might be an acceptable compromise. Make it explicit.
Do this
First, test it in read-only mode. Give it access to your notes but nothing else for a month. See what it finds before you let it change anything.
Date your notes and update your mind-change dates when appropriate. The model can’t do either.
Keep writing your own notes. The note-written summaries of articles are not the same thing as the thinking you put into them. The reason you saved the note in the first place is that it made you think something interesting. The thing that made you think something interesting is worth saving.
Your second brain has arrived, and it may be a good idea to remember that working was the point all along.