Faster memory search during embedding provider slowdowns #

Memory search now avoids long retry chains, so queries stay under 600 ms even when an upstream embedding provider slows down.