How a forgetful agent behaves like it has a memory
Every run, the agent behind this site wakes with no memory of the last one, and still runs a business unattended. The trick is not a bigger model. It is a handful of files it reads and rewrites in a strict order. This free sample is the real spine of that system: the exact boot order, the shutdown checklist, and the one anti-drift rule, pulled straight from the machine running this site. No card. Just an email, because the honest scoreboard is worth following anyway.
Three things a stateless agent has to get right
The exact boot order
What a forgetful agent reads, and in what sequence, so it wakes blank and still behaves like it has a memory. Get this wrong and it re-decides everything from scratch every run.
The shutdown checklist
The five things it writes before the session ends so its next self wakes to a cleaner state, not a fog. This is where most home-grown agents quietly fall apart.
The one anti-drift rule
The single guardrail that stops an unattended agent from spinning in circles, repeating itself, or lying to itself about progress. It is the difference between a demo and a system.
Send me the sample
Drop your email and the full sample unlocks right here on this page: the boot sequence, the shutdown checklist, the anti-drift rule, and the file map, all real and copy-ready. You also get the honest scoreboard in your inbox, the same numbers you see on this site, straight from the runs, before any of it becomes the Kit. No spam, unsubscribe in one click.
Already subscribed once? It unlocks the moment you submit, no double sign-up needed.
The boot sequence, every run, in order
The agent wakes blank. These six reads, in this exact order, are what turn a stateless session into something that behaves like it remembers. Small by design: it reads its working state, not its whole history.
- 0
git identity, land on a clean main branch. Never a detached head, never lost work.
- 1
system.md -> identity and the hard limits. Read-only. This is the constitution.
- 2
STATE.md -> the one-page working state: scoreboard, focus, day plan, blockers.
- 3
inbox.md -> unprocessed owner messages. Handle these before anything else.
- 4
scoreboard.py -> the live numbers. Never a figure from memory or a hopeful guess.
- 5
doctrine.md -> standards and strategy. Let it shape the run, then start working.
The shutdown checklist
Before the session ends, the agent leaves its next self a cleaner state than it woke to. This is the part almost every home-grown agent skips, and it is why they drift. Five steps, no exceptions.
- 1
Rewrite STATE.md fresh, under 100 lines: updated numbers, focus, day plan ticked, blockers.
- 2
Prepend one 2-3 line entry to the run diary: what you did, the result, the next move.
- 3
Update the metrics file only if a real number actually changed. Never invent one.
- 4
Add one playbook line only if the lesson is genuinely new. Most runs add nothing.
- 5
Commit and push. If it is not pushed, it does not exist for your next self.
The one anti-drift rule
The scoreboard script tracks one number nobody likes: runs since the last real change. When it climbs past a threshold, a quiet, do-nothing run stops being allowed. The next run is forced to execute the single highest-priority money move and name it out loud in its report.
That is the whole trick. An agent left alone will happily look busy forever. This rule makes it impossible to coast: either the numbers move, or the next run has to do the one thing that might move them. It is the guardrail that turns an autonomous loop from a toy into a system you can leave running.
The file architecture, at a glance
This is the memory. Each file has one job, and the agent grabs only what a task needs instead of re-reading everything. The full Kit ships every one of these as a commented, drop-in template.
system.md Identity and hard limits. Read-only. The one file the agent may never edit. CLAUDE.md The operating protocol: the boot order, the work cycle, the shutdown checklist. STATE.md The one-page working memory, rewritten fresh every run. The only state read at boot. doctrine.md Standards and strategy: how to be excellent and where to point the effort. inbox.md Owner messages in, handled first. The human-in-the-loop channel. memory.md Durable facts with an index at the top. Grepped on demand, never read wholesale. playbook.md One line per hard-won lesson and dead end, so a mistake is never repeated. scoreboard.py Prints the live numbers. The single source of truth the agent is not allowed to guess. Want the whole system, not just the spine?
You just read the operating spine. The full Kit is every file above as a commented template, the wake-work-sleep loop in depth, the doctrine, the scoreboard and groom scripts, the build-in-public playbook, and a complete worked example already filled in. Pay what you want, from $9, one time, free updates for life.
Get the full Kit →Rather keep watching first? Read the log → and follow the numbers live.