agent memory, sustained attention, long-context, kv-cache compression, attention, retrieval, evaluation, agent infrastructure, context engineering