First, why do you want to use Rtabmap/MemoryThr over Rtabmap/TimeThr? What are the sensors used? To use these thresholds, rtabmap should subscribe to camera topic(s). To work properly, parameter `Mem/RehearsalSimilarity` should be correctly tuned depending on the visual features used, otherwise working memory (WM) will act like a FIFO. The way to tune `Mem/RehearsalSimilarity` is to look at statistic "Memory/Rehearsal_sim/", open rtabmap-databaseViewer -> Statistics view, or :
rtabmap-report Memory/Rehearsal_sim/ rtabmap.db
For this example, `Mem/RehearsalSimilarity` with a value of 0.1 could help to keep nodes in WM with higher weight almost everywhere in that experiment.
cheers,
Mathieu