What we're currently trying to do is to create a new option for append mode such that it will only record poses after loop closure detection.
It should be already doing that. The new frames won't be recorded in database till the first localization happens with previous session. See from 4:15 in the
video.
With Mem/LocalizationDataSaved=true, if you start a localization session, rtabmap will still save the received frames in the database but they will be disjoint from the original session. We can browse them in rtabmap-databaseViewer, but they won't be linked to map in that database.
cheers,
Mathieu