# Continual learning

Improve over time while keeping changes observable.

Source: https://docs.minds.sh/docs/concepts/continual-learning



Continual learning is the process of updating a system as new experience arrives. Instead of treating a model or learned state as permanently fixed, an application can run training tasks, inspect their progress, and compare the resulting behavior.

It is different from remembering a document. Storing a new record makes information available; training changes the behavior of a learned component. Choose the operation that matches your goal.

## Make a change measurable [#make-a-change-measurable]

Define the task, inputs, expected improvement, and evaluation before starting a learning job. Keep a stable comparison set so a new result can be compared with the previous one. Record configuration and model identity alongside the outcome.

A useful evaluation asks both whether the new task improved and whether older behavior regressed. A lower training loss alone may not answer either question for your application.

## Recovery must be real [#recovery-must-be-real]

Checkpoints and rollback are valuable only when they actually save and restore the relevant state. The current daemon's checkpoint-status and recovery handlers acknowledge requests without implementing restoration, and no checkpoint-creation HTTP route is registered in the audited snapshot. Do not design a production recovery procedure around those acknowledgments.

Use the [continual API](/docs/api/instance/continual-learning) for available task contracts and [backup guidance](/docs/operations/backups) to distinguish data backups from model or training checkpoints.