Less work_mem because this explodes easily. It's per-operation, which a
query can have multiple of, and also across multiple worker threads.

More shared_buffers and effective_cache_size as these are global.