Docs / What 1.x promises

What 1.x promises

A 1.0 version number means more than a feature list. The promises below hold for every 1.x release, and each one has a check behind it: automated tests for the store format and the tools, an install over the previous release before each release ships, and a recorded check that a 1.0 store opens in 0.6.8. A release that breaks one of them won't be called 1.x.

Your store keeps opening

A store written by any 0.5.x or 0.6.x release opens in every 1.x release as it is, with no re-index and no migration step. The on-disk formats and the way notes are split into pieces are frozen for 1.x. A later 1.x may learn to read more, but will never stop reading what you have now. Every release is tested against small stores written by the code of 0.5.1, 0.6.0, 0.6.6 and 0.6.8, including 0.6 stores that have been compacted and checkpointed.

Opening a store only to read it changes none of its files. If a 1.x release finds data written by a newer ModelBrain, it says so and stops, and changes nothing on disk.

You can go back

1.0 writes no new format, so a store used by 1.0 still opens in 0.6.8. This was checked by opening a store written by the 1.0 code with 0.6.8. Going back to 0.5.x is not supported.

Tools and settings keep their meaning

The 14 tools your assistants call (remember, recall, list_memories, expand, forget, the six partition_ tools, diagnostics, ping and usage) keep their names, their required fields, their documented response fields and their error codes through 1.x. New fields are always optional.

These settings keep their meaning through 1.x. Any other MODELBRAIN_ name, including ones in older release notes, is internal and may change.

Upgrades are quiet

Installing 1.x over 0.6.x or an earlier 1.x stops ModelBrain and starts it again, and opens your store as it was. It leaves your assistants' settings and tokens alone, doesn't download the reasoning model again, and re-indexes nothing. Before each release ships, it is installed over the previous one on macOS, on Windows on Arm and on Linux. The Windows x64 installer is checked on an x64 PC after each release.

Saves reach the disk

ModelBrain writes each save to disk before it replies. A crash or power cut loses at most the save that was in progress, and ModelBrain tells you if it finds anything already saved that it can't read. If the disk fills up, saves fail with storage_full while recall keeps working, and saves start working again by themselves once there is space. In the rare case that doesn't work, the log says to restart ModelBrain. An encrypted snapshot is taken once a day by default and the last seven are kept. See Local backups and restore.

No network call you didn't ask for

ModelBrain makes no network call unless you turn something on. Turning on Use Reasoning Mode in the tray downloads the reasoning model once. The ChatGPT or Grok relay, sync and managed backup need an account you create. There is no update check and no telemetry.

What these promises don't cover

ModelBrain decides which assistants it serves and what each one may delete. That doesn't stop a program that is already running as you. Your store isn't encrypted at rest on your own disk. The Windows installers are signed by NexSpark LLC, but SmartScreen may still warn for a while after each release until enough people have downloaded it.

Release history

What changed in each version.

Changelog →
Backups

The daily snapshot and how to restore.

Local backups →