Makes sense to reduce disk accesses but maybe there are fixed data structures allocated? Since otherwise they should advice a given percentage of memory compared to dataset size instead of a fixed amount.

You may find this article helpful: http://rethinkdb.com/docs/memory-usage/

The article describes how RethinkDB uses memory to reduce disk access and process queries, as well as how to reduce its memory usage with `--cache-size`.

Because it's 2015.