...Its Memory Parallel engine distributes scoring and transfers only selected memory back to the accelerator. Memory Interleave alternates retrieval, context expansion, and generation to improve multi-hop reasoning across distant segments. The project reports experiments extending from 16K to 100M tokens, including inference on two A800 GPUs.