Happy to share this and curious how others are thinking about runtime-level architectures for LLM cognition.