Memory architecture for AI teams
Your AI gets less reliable
the more your
team uses it.
Because every session
starts cold.
Most teams answer AI drift with a bigger model or a longer context window. That is the wrong layer. The teams pulling ahead wrote the brain down, and they check that it got read. Here is the whole architecture, open on the page.
The argument
A smarter model with no memory is a smarter intern with amnesia.
Every new AI session opens like day one on the job. Smart, fast, confident, and completely unaware of anything the team decided last week. Nothing important is written down in a form the machine reads before it acts. That is the whole problem.
Bigger models and longer context windows do not touch it. This is not a model problem. It is an architecture problem, and it is fast to fix once you name it.
The three files
The Written Brain is three plain files the AI reads before it works.
Copyable, editable in any text editor, versionable in git. Nothing exotic. The discipline is that the machine reads them every session and writes back the moment a call gets made.
What is true right now.
The current position of the work. Who owns what, what is shipping, what is blocked, and what changed since last session. The living top-of-mind for anyone (or anything) opening the project.
Every call we made, and why.
Locked decisions with their reasoning. This is the file that stops the AI from re-opening settled calls in a new session. When a question comes up that is already decided, the AI points at the line instead of relitigating.
Facts still true, with source.
Verified facts with the file, thread, or source they came from. Contested items live separately. The AI grounds its work in this file before it invents anything.
The protocol
Bootkers & Bonkers.
The dumbest word in your stack, and the reason it works.
A written brain only works if the machine reads it every time. Documentation nobody reads is not architecture, it is decoration. The protocol makes the read enforceable.
Session start
Before the AI touches any work, it has to type one word: Bootkers. Yes, made up. The dumber the word the better, because nobody types Bootkers by accident. If you see it, the files got read. If you do not, the AI does not get to work.
Session end
When the session ends, the AI has to type a second one: Bonkers. That means every decision made this session got written back to decisions.md. No Bonkers, the job is not done, and yes, you check.
Pick your absurd word. You will thank me later.
What it changes
Consistency that survives scale.
New agent or new hire, productive in a morning.
The context is written down instead of stuck in five people's heads. A fresh AI session is doing real work in minutes. A new person could be doing it before lunch.
Every decision, traceable.
When the board asks how a call got made, you point at the line in decisions.md with the date and the reasoning. Instead of hoping.
The open questions never close on their own.
The AI surfaces the call and waits for a person. That one rule is the difference between automation you trust and automation you babysit.
Six weeks in, the work gets better.
Every session adds to the brain instead of forgetting it. What used to feel like starting over becomes stacking. The pilot becomes the system.
Built with you
The Written Brain
Setup Sprint.
Five days,
$4,500,
yours forever.
The framework on this page is free. Take it, build it yourself. If you want it built with you, on your existing AI stack, tuned to your team's actual work, that is the Setup Sprint.
What ships in five days.
- The three files (state.md, decisions.md, knowledge.md), populated with your project's real content, not templates.
- The Bootkers & Bonkers protocol wired into your Claude, ChatGPT, or Gemini instance, with the enforceable read.
- One narrow-agent split from an existing over-scoped agent on your stack.
- A two-hour handoff session so your team owns the setup, not me.
- The absurd word is your call.
Send this to someone
If someone on your team keeps rebuilding context every Monday.
Copy the link. Send it to them. The framework is on the page, no download, no gate. Also lives at jeffschenck.com/written-brain.
Start
If your AI keeps
re-explaining itself,
you have an
architecture problem.
Bring the setup. Leave with a real plan.