Build With a Memory Is Now a Public Repo

July 17, 2026

I wrote about giving the agent a memory before you give it a task. That post was the what. This is the how.

The whole method is now public: buildwithamemory.com, backed by a repo you can read, copy, and argue with at github.com/adetwiler/build-with-a-memory.

What’s actually in it

Twelve short method docs. Copyable templates. A one-paste prompt that sets a project up from zero. All plain markdown.

There is nothing to install and nothing to sign up for. That’s deliberate. Your project’s context should live in your repo, as files you own, next to the code it describes. Not in somebody’s hosted memory product. I would rather show you the method and have you keep the files than sell you a subscription to your own context.

It works with whatever agent you use. The files are just markdown; any tool that can read your repo can read them.

The part I’d start with

If you only take one thing, take the decision records.

When you make a real decision, write one small file: what you decided, and why. The why is the part you forget first.

Here’s what that buys you. Normally, review cost grows with a project. More code, more history, more “wait, why is it like this?” With decision records, it inverts. The agent reads the decision once and stops re-litigating it. Every session after that is cheaper than the last. Two months in, my reviews got faster, not slower.

That inversion is the whole reason the rest of the method pays off.

What it won’t do

Honestly, it’s not magic. The agent still gets things wrong. The files go stale if you stop tending them. And a pile of docs nobody maintains is worse than one page that’s true. The method is mostly discipline: write the decision down when you make it, write the gotcha down when it bites you, keep the always-loaded page short.

I’m still figuring out parts of it. The repo says so where that’s true.

Where this came from

I didn’t sit down to design a method. It fell out of building real things: my games, and Playloop, the playtest-insights tool I’m building for indie devs. Every gotcha in those repos taught me something about what an agent needs written down. The private projects keep their private files. The method is what generalizes.

If you want to follow along as it grows, I write a short letter when there’s something worth saying. And if the playtesting side of this sounds like your problem too, Playloop is coming later this year and has a waitlist.

But mostly: go take the files. That’s what they’re for. If you try this on a real project and something doesn’t hold up, tell me. The method got better every time something broke, and I don’t expect that to stop.

Credits

Not all of this is mine. The mindset shift came from a talk by Matt Pocock, and the grilling pillar adapts his grill-with-docs skill. The repo names every borrowed idea in CREDITS.md. If I borrowed something without naming it, tell me and I will fix it there.

How I Built a Local Memory Network for My AI Agents
No hosted database, no SaaS. Markdown as the source of truth, links that make it navigable, hooks that surface it, and a disposable local index for search.
Give the Agent a Memory Before You Give It a Task
AI agents start every session at zero. The fix isn't a smarter model. It's writing things down where the work happens.