oMLX uses a two-tier RAM + SSD KV cache architecture that keeps active entries in RAM and offloads inactive ones to SSD, with the cache persisting across application restarts.