One install, three MCPs: graph chat memory + semantic code search + backend. Benchmarked with claude-haiku-4-5: -19% tokens & 23× faster on memory-recall questions.