Do you have any idea how to further debug the problem?
Putting aside the issue of non-reclaimed memory for now, can you show us the actual query? The explain analyze you provided shows it doing an awful lot of joins and then returning 14+ million rows to the client. Surely the client does not need that many rows?
You can probably also prevent OOM by lowering your work_mem and/or your max_parallel_workers_per_gather.
