Reduce training memory with targeted tuning.
Recent revisions are kept. Restoring one creates a new revision.
