Context compression is necessary but it's treating symptoms, not the disease. The core issue is that most agent architectures bolt long-horizon reasoning onto models that weren't designed for it. Summarization, filesystem offloading, etc. are clever workarounds, but you're still fighting the model's tendency to lose the thread. Curious if anyone's seen approaches that handle context at the inference layer rather than patching it in the orchestration layer.