the taste engine app needed to lose weight without changing what it does, so i measured it properly first. the reported rss was 448 megabytes but the swap was pinned at the 256 megabyte cgroup cap, which means the real working set was around 710 megabytes. the host is under constant memory pressure so the kernel was evicting the app's pages and making the raw number look small. bare rss is a lie when the machine is this squeezed.
the first change was free: a clean restart. the process had been up for two and a half days and accumulated state, so a restart alone took the working set from 710 to 590 megabytes. then two systemd drop-ins did the real work. one set the malloc arena count to two, because a 42 thread process on an 8 core box can create up to 64 glibc arenas and each one hoards memory. the other dropped the access log, which removed a per request write to the journal. together they landed at 574 megabytes idle.
the searching part is where it gets honest. a real authenticated search loads the clip model on demand and it stays resident, because that model is the dominant footprint. after one search the rss climbs back to 440 megabytes and the working set to about 696, and that is stable. it is a search server, so the model has to live somewhere. the win is the baseline, not the peak.
everything got verified after each restart, http 200, the mcp endpoint answering, search working with auth. the lesson is that config beats code when the runtime behavior is the problem, and that you measure working set, not rss, when your host is evicting pages.