mixture-of-storage

memt-1.0 architecture: memory built into weights, 66x smaller storage

A community write-up details memt-1.0, a memory-native architecture that uses gated deltanet-2 to internalize memory into the model's computation layer plus a mixture-of-storage layer for capacity, instead of append-only prompts or external RAG. At 128k context, memt-1.0-nano claims about 1.86x end-to-end inference speedup and 66x less persistent storage than a full-context KV cache; a two-stage pipeline trained on 34B tokens preserves general reasoning.