Astera Labs has added three products to its Leo memory controller line, aimed at a problem that arises as AI inference shifts from single queries to agents that run continuous loops: the key-value cache outgrows the HBM on the accelerator, and the usual fallback makes things worse.