I replaced Notion, bookmarks and my memory with one folder
A few weeks ago, sometime past midnight, I caught myself asking an AI how to write a zsh helper I had already solved. Twice. Both answers were sitting somewhere in months of chat history, and I could not dig up either one.
I closed the laptop and did honest math on where my knowledge actually lived. A bookmarks bar I had not opened in forever. A Notion workspace I spent a whole weekend decorating, then visited maybe four times, total. Forty tabs open as a to-do list.
The pattern embarrassed me, because capturing was never my problem. Keeping things was. Every system I tried was great on day one and a landfill by month two.
So I did the laziest thing possible.
The night it started
One evening I opened a terminal and typed mkdir brain. That is the entire migration story. No app comparison spreadsheet, no productivity YouTube, nothing. Three files, badly written, and one of them was literally called stuff.md.
What sold me on plain markdown before I even started: every tool I ever loved got acquired, paywalled, or shut down eventually. Plain text has outlived all of them. I wanted my memory stored somewhere no company could take hostage, and that place is a folder.
The rules came from breaking things
Nothing about the layout was designed. All of it is scar tissue.
Everything lived in one file at first. Around line 300 I could not find anything inside my own brain anymore, which is how I learned that one file should equal one domain. Tables looked organized and took effort to maintain, so bullet lists won. A file crossing roughly 200 lines became impossible to skim, so that is the split point now.
Four folders appeared on their own along the way:
prompts/— reusable prompts grouped by use-caseresources/— tools, links, references (this replaced my bookmarks)chats/summaries/— takeaways distilled from AI chatsimports/— raw dumps waiting to be processed
More folders grew later, whenever one topic kept polluting a file that had better things to do. Mine has a couple I will not describe in a public post. If yours grows differently, that probably means yours is working.
The imports folder saved the whole system, honestly. Older setups of mine died because dumping something in felt like creating homework. An official junk drawer fixed that. Ten seconds, dump, close laptop, process later or never.
Then I pointed an agent at it
For the first few weeks it was just a tidy folder. Nice, but a folder.
Then one day I started a completely fresh chat with a terminal agent and gave it exactly one instruction: read ~/brain. And it walked in knowing me. Which tools I run. A project I had dropped in June, including the reason I dropped it, because past-me had documented the reason in a file I had forgotten I wrote. I did not explain myself once. It did not ask me a single setup question.
I sat back for a second after that one.
Since then it keeps doing small things that feel like cheating. I will ask about some task and hear back that I already wrote a prompt for it. I will start researching a tool and get reminded that I evaluated the same tool two months ago and moved on, here are my own recorded reasons. My memory leaks constantly. Apparently the leaks have been landing in a folder this whole time, and now something reads them back before I repeat myself.
So this is the real use case, the thing a pile of notes could never do. The brain gives every agent I work with full context from minute zero. Tiny notes matter most here, the half-formed embarrassing ones especially, because those are the connections that resurface later. The way I grade the system now: fresh chat, hand over the folder, count the questions it asks me. Close to zero means the brain is complete. When it is not zero, that is a missing file, not a bad conversation. Fixable in one evening.
What using it actually looks like
Early versions of my workflow had me dictating paths. Summarize this into that file. Add this tool under that heading. It worked, but something kept bugging me: if I have to tell my own brain where to put things, is it really a brain?
The fix was one more plain markdown file. A rulebook, sitting at the root of the folder. Conventions, formats, where each kind of thing belongs, what should happen after certain kinds of sessions. Every agent pointed at the folder reads that file before anything else, so the brain ships with its own operating manual. New agent, same behavior, zero explanation from me. That one file is why the back-and-forth disappeared.
Now capturing looks like this:
me: look into this later <link>
agent: read it, it is a CLI tool. Filed under resources/tools.md,
next to your other terminal stuff.
Notice who said the path. Not me. I gave it four lazy words and it matched the link against the existing structure and picked the spot. Some days I genuinely do not know which file something landed in, and I have stopped needing to know.
Recall turned out to be the bigger upgrade anyway. Mid-task, I type things like:
wait, didn't I already research message queues?
and get my own conclusions back from weeks ago. My words, my tradeoffs, my reasons for moving on. Not a random article, my notes. The first few times that happened it felt genuinely strange, like the folder was finishing my sentences.
Starting anything new goes through the same door. Whatever the task is, the agent is reading it next to everything similar I have already done: old decisions, prompts I reuse, approaches I tried and abandoned with the reason recorded. Doing the same work twice became almost hard to do by accident.
Filing was the reason I quit every notes app I ever picked up. These days filing is not even my job.
Viewing and syncing cost nothing
A friend asked what editor this whole thing needs. None, really. Opening the folder as an Obsidian vault gives graph view and backlinks if I want them, VS Code works, sometimes I edit straight in the terminal. Text does not care where it gets typed.
Sync is git:
cd ~/brain && git init
gh repo create brain --private --source=. --push
Version history, sync between machines, no new service to trust.
Why nobody gets the folder
Someone asked if I would share the whole thing recently. Hard no. There are half-finished ideas in there, blunt self-reviews, notes to myself I would not want read by anyone. So the line I drew early on holds: the system is public, the contents mostly are not. A brain curated for an audience slowly stops telling you the truth, and then there was no point building it.
What I would tell a friend
Skip my folders, steal the constraints:
- plain markdown, nothing proprietary
- one file per domain, split past ~200 lines
- a dump folder, so capturing costs nothing
- summaries over transcripts
- let an agent read it and file into it, and grade yourself by how few questions it asks you
- a rulebook file at the root, so agents inherit the system instead of being dictated paths
Mine began as three badly written files, one named stuff.md, and it worked anyway. The system does not need to be impressive. It needs to catch what your head drops.
Mine catches almost everything now. That zsh question never came back.
