I would be opposed to this, since there's no support for Valgrind since jemalloc version 5.0.
As an example rationale: LibreSSL was forked from OpenSSL because of OpenSSL's custom memory allocator and the OpenBSD project not catching memory safety issues because of it. (It wasn't because of Heartbleed.) I want to rely on libc's malloc(3) and would not appreciate an additional dependency. Smart ideas may not always be smart in the end. I think you'd be better off convincing to switch the malloc(3) system call to use jemalloc(3) everywhere, if the benefits are so good as claimed. You can do you by adjusting the variables in your environment. Also, your memory issues are probably from big rooms like #matrix:matrix.org with lots of federating servers and 4.3K+ users; that's expected of being there for a long time. For what it's worth, joining that room is doable on a 512 MB virtual machine just fine with memory consumption < 100 MB (at least for a hour or few with one user). I also don't think this bug warrants "normal" severity; likely "wishlist", maybe "minor".
