It seems we are not including the 8 bytes per node pointer in the analysis. Apparently there has been a recent push to improve the accuracy of memory tracking.

If we decide to add it to the tracking, we should also consider if we want to adjust the spilling threshold. Otherwise, we would be spilling slightly more frequently, with fewer rows at a time.

http://10.1.2.40/mongodb/source/src/mongo/db/pipeline/group_processor.cpp?q=%2Bfunction%3Amongo%3A%3AGroupProcessor%3A%3AshouldSpillWithAttemptToSaveMemory%28%29&redirect_type=single#215

http://10.1.2.40/mongodb/source/src/mongo/db/pipeline/group_processor_base.cpp#168 

See a similar ticket for $graphLookup https://jira.mongodb.org/browse/SERVER-90439