Can an LLM Forget the Right Things?

Standard LLM inference runtimes are ill-suited for real-time robotics, as they overflow memory with endless video streams and silently miss critical physical deadlines. A custom-built CUDA runtime solves this with an admission controller that refuses to start any reasoning task it cannot finish within a strict 3