Generally, as you may know, it is not very easy to get memory usage stats on Unix/Linux without specialised tooling.
However, in my experience, with default TCP settings, the kernel overhead plus whatever "typical" app overhead will not often exceed say 100 kB per connection.
Now, we had originally used an even smaller 1GB server than yours, and we did run out of memory, with attendant catastrophic failures. We were able to trace the failures to certain ORM queries on tables with large JSONB columns (in the MB range). Once we addressed that, we were fine. At no point, even with our usage pattern of connect per message (!!!) , did sockets usage even warrant a second thought.