Threads have always felt limited—long conversations don't fit neatly into projects, and splitting them across multiple threads loses context. Search + labels might be a better model. What if we treated conversations with machines more like debug logs? The LLM could tag sections in the background as it manages its context window, making it easier to surface patterns, flag key moments, or parse the conversation history later.