How it fixes itself
How the device checks itself and fixes small problems on its own.
It aims to check the device every 90 seconds: its services, the AI, memory, disk, temperature, network, voice and screen. If a check finds a real fault, it tries one small, fixed repair, such as restarting the assistant. It writes down every repair and the reason for it.
Good to know
- Every repair has a limit. For example, one service is restarted at most three times in half an hour.
- Repairs that change files can be undone. If one did not work, the device puts things back as they were.
- It remembers which repairs worked on this unit and prefers those.
- When the disk runs low, it trims its own logs and caches. If the disk is critically full, it can delete older backups, but it keeps at least the two newest.
- As a last resort it restarts the whole device, at most once a day. It tells you first, with how long until the restart. It never does this in the first 30 minutes after start-up or while an answer is being written.
- It never changes its own software and never installs updates by itself. It never touches your messages, notes or reminders.
- Tools for testing and improving the device (idle self-analysis, practice faults and code experiments) are switched off on shipped units.
- To see what it fixed, tap Device. Overview lists the latest repairs; Activity shows What the device fixed by itself.
If something goes wrong
| What you see | What to do |
|---|---|
| “I’ve stopped retrying, see System Health” | Tap Help → Open Device and read what it says. Restart the device, or contact [email protected]. |
| The device says it will restart itself | Let it. It saves a record of the problem first. |
| It asks you to unplug it and plug it back in | Today’s one automatic restart was already used. Unplug the power, then plug it back in. |
Privacy
Self-repair runs on the device. It never goes online and never touches your accounts.
Related topics
More detail
The testing tools (off on shipped units)
Idle analysis by the local model is off by default. Deliberate fault rehearsals are off by default. Proposals for new engineering work and code-change experiments are off by default.
This self-check does not search your conversation history or saved memories. It reads diagnostic records, which may contain text recorded with a fault. It yields when new activity is detected; stopping it or preparing the next answer can still cause a delay.
The repairs
Eighteen repairs aim at a fault. Beyond them, the device only tidies up after itself.
| repair | what it fixes | can be undone? | limit |
|---|---|---|---|
| Restart the assistant | the chat service stopped answering | no (a restart) | 3 per 30 min |
| Restart the model server | the AI server stopped answering, or grew too large | no | 3 per 30 min |
| Restart the Telegram bridge | the Telegram connection went down | no | 3 per 30 min |
| Re-warm the models | an AI model dropped out of memory | n/a (a reload) | as memory allows |
| Relieve memory pressure | memory critically low twice in a row | no | 3 per 30 min |
| Restart the face | the screen stopped drawing | no | 3 per 30 min |
| Re-run the boot self-test | the start-up check had not passed | n/a (a check) | 3 per 30 min |
| Rejoin Wi-Fi | the device is off a network it knows | no (changes no password) | 3 per 30 min |
| Restore a settings file | a shipped settings file changed from its known copy | yes: checked, undone if wrong | 3 per 6 h, 2 undos a day |
| Rebuild the device facts file | the file of non-secret device facts changed | yes | 3 per 6 h, 2 undos a day |
| Repair the memory index | the search index disagrees with your memories | yes | 3 per 6 h, 2 undos a day |
| Rebuild the memory index | the search index is too damaged to repair | yes | 1 a day (counts toward the 3 per 6 h), 2 undos a day |
| Repair the conversation index | the conversation index disagrees with your kept messages | yes | 3 per 6 h, 2 undos a day |
| Rebuild the conversation index | the conversation index is too damaged to repair | yes | 1 a day (counts toward the 3 per 6 h), 2 undos a day |
| Rebuild its own history | a history the device builds for itself no longer matches its record | yes | 3 per 6 h, 2 undos a day |
| Restart the face overnight | the screen’s browser uses too much memory (between 3 and 5 in the morning, after ten quiet minutes) | no | hourly checks; shares the face-restart limit |
| Reload the model runner | the chat model was pushed out to slow storage by other work | n/a (a reload) | 1 per hour, only while idle |
| Restart the device (last resort) | the AI, the chat service or the background service has been down for 30 minutes and its repairs are used up. You are told as soon as a repair fails, with how long until the restart, and told again when it works. Never in the first 30 minutes after start-up, during the device’s own test windows, or while an answer is being written. Once you have been warned, using the screen no longer delays it. If today’s one restart was already used, you are asked to unplug and replug it instead. | no (a reboot; a record is saved first) | 1 a day |
After an undo, the device does not retry that repair for six hours. It can also tell when a fault is a cause and when it is only a result, and it says so when it cannot tell. A repair aimed at a symptom is held back when the evidence points elsewhere. A change to the software reaches the device only as an update you install: the device checks the signature and every file first, and attempts a verified rollback if its self-test fails.
Self-repair is not the Autonomous AI mode. That is a separate choice; see AI modes.
From the device manual for software 0.9.0 (November 15, 2026). Your device serves the manual for its installed software; where this copy and your device differ, your device is right.