Your AI agent starts every session knowing nothing about your project.
I've already published the full internals of my memory network. Fair warning applies to that post: it's months of accretion across every project I run, and nobody should start there.
So this post is the other end of it. The five-minute version. One paste, and what it actually produced on a real repo.
I ran the setup prompt on this site's repository. The repo behind andrewdetwiler.com. It felt right to test it on something you can see.
The state before the run: 34 commits, a README, no memory files for the agent. An agent opening this repo cold would know nothing about the em dash gate, the embargo list, or why the short links look the way they do.
I pasted the prompt into a fresh agent session and let it work.
Five small files, about ninety lines total:
CLAUDE.md
, the main file. What the site is, the real commands, and the rules that matter. It pulled the rules out of the README instead of inventing them: the em dash gate, the everything-is-ungated rule, the near-zero JS posture.AGENTS.md
, a two-line pointer at the main file, so any tool finds it.docs/now.md
, what's active right now. It read the recent commits and wrote down that a blog post was in wording passes.docs/decisions.md
, seeded with four decisions it found in the history. My favorite: the embargo word list is kept out of git on purpose, because the list itself would leak the names. That reasoning was buried in the README. Now it's a dated decision the agent reads in one line.docs/notes.md
, three gotchas it noticed just by reading. One of them was real and I had never written it down: the pre-commit hook does nothing until you run git config core.hooksPath .githooks
in your clone. A fresh clone commits right past the gate.That last one is the part worth pausing on. Nobody told it that gotcha. It's in the repo, if you read carefully. I hadn't written it anywhere, so every fresh session was one careless commit away from learning it the hard way.
Honest scope, because this is where prompt packs usually oversell.
It didn't overwrite anything. This repo already had two context files (a glossary and an open-items list), and the prompt is written to add around what exists, not replace it. It read them and pointed at them from the main file.
It didn't produce deep knowledge. Ninety lines from a cold read is orientation, not understanding. The decisions file starts with four entries because that's what the history could support. The real value shows up later, when the habit section it wires in starts catching decisions the day you make them.
And it's a prompt, not a program. The agent follows instructions, so you check the short report it hands back. Mine got this repo right, but this repo is small and the README was honest. Yours may need a correction pass.
The prompt is free and ungated at buildwithamemory.com, with the raw text in the repo if you want it without the page around it.
It works with whatever agent you use. The output is plain markdown in your repo, files you own, next to the code they describe.
Run it once, then the habit does the rest: when you make a decision, it gets a dated line and a reason. When something bites you, it gets a line right then. Next session, the agent reads those first.
If you run it, I'd honestly love to know what it wrote and where it missed. The prompt is written to say so when a repo is empty or messy, but I've only got my own repos to test against, and they all have my habits baked in. The prompt and the demo run are my own work. The method behind them borrows from named people, and the record lives in the repo: CREDITS.md.