Step 10 — Backups
Step 10 · Honesty
Everything the agent knows lives on one computer in your house. Losing that machine should be an inconvenience, not the end of the project.
The trap is that the things worth backing up are not the things that take up the most space.
Two backups, for two problems
Section titled “Two backups, for two problems”| Recovery backup | Convenience backup | |
|---|---|---|
| Purpose | Survive losing the machine | Undo a mistake from earlier today |
| Contains | Only irreplaceable state | Everything |
| Size | Small | Large |
| Where | Off the machine | On the machine |
| Frequency | Daily, automatic | Manual, when you want one |
You need both, and confusing them causes trouble. A large convenient backup is not disaster recovery, and a small recovery backup will not roll back an hour.
Do this
Section titled “Do this”- List what cannot be rebuilt. That list is your backup.
- Back it up daily, automatically, off the machine.
- Add a size check so a broken run cannot overwrite a good copy.
- Restore it once, on purpose, before you need it.
- Write down what is not covered.
What goes in, and what stays out
Section titled “What goes in, and what stays out”The principle: back up what cannot be rebuilt. Everything else makes the backup slow and the restore fragile.
| Back up | Why |
|---|---|
| Working memory and conversation history | Not reproducible, and it is the whole point |
| The wiki | Your household’s knowledge, written by hand |
| Configuration and schedules | Rebuilding by hand is error-prone |
| Credentials, encrypted | Painful to re-issue |
| The job definitions | Reproducible, but tedious |
| Leave out | Why |
|---|---|
| Downloaded programs and dependencies | Reinstallable in minutes |
| Caches and temporary files | Regenerate themselves |
| Log files | Useful for a day, then noise |
| Large media attachments | The conversation survives without them |
That second table is what takes a backup from gigabytes to something you can email. Mine went from roughly two gigabytes to tens of megabytes, which is what makes keeping an off-site copy practical.
Make it fail safely
Section titled “Make it fail safely”The most dangerous backup is one that silently writes a broken copy over the last good one. Two habits prevent it:
- Write to a temporary file, check it, then move it into place. If the job dies halfway, the last good copy is untouched.
- Refuse to publish a suspiciously small archive. A size check catches the common failures: a disk filling up, a half-finished run, a folder that went missing.
A backup you have never restored is a hypothesis. Restore it once, deliberately, onto a spare machine, before you need to.
Restoring onto a new machine
Section titled “Restoring onto a new machine”Roughly an hour, once the downloads finish.
- Install the operating system and the agent platform.
- Create the agent’s own account again — a clean machine will not have it.
- Restore the recovery backup.
- Re-authenticate every external service. Credentials rarely survive a machine move, and the failures are confusing if you do not expect them.
- Re-pair every channel the household talks through.
- Confirm the schedules are back, and send yourself a test message.
Step 4 is the one that catches people. Access to external services is usually tied to the machine in a way that does not transfer, and no amount of restoring fixes it. Budget for it.
The gap I have not closed
Section titled “The gap I have not closed”A daily recovery backup has one honest weakness: if the state is corrupted today, tonight’s backup faithfully overwrites the good copy with the corrupted one.
You are protected against losing the machine. You are not protected against realising at 4pm that you needed yesterday’s version.
My mitigations, cheapest first:
- A second, local snapshot that keeps hourly history for a day or so. The cheapest real fix.
- Rotate a small number of daily copies — a week, not a year — if storage allows.
- Accept it, and write it down. An acknowledged gap is survivable.
It is on The Gaps I Have Not Solved deliberately. I would rather tell you than have you find out.
Checkpoint
Section titled “Checkpoint”- A daily backup runs and you have seen today’s file
- It lives somewhere other than the agent’s machine
- A size check would stop a broken copy replacing a good one
- You have restored it once, on purpose
- You have written down what it does not cover
Watchdogs — because a backup that stopped running three weeks ago is the most common silent failure there is.