Context that Compounds: Memory, Learning + Long-Running Agents

Long-running agents generate more context than they can keep in the prompt, so the hard part is deciding what survives. Josh Bleecher Snyder of exe.dev leads Caitlin Colgrove of Hex, Tyler Edwards of Overmind, and Pasha Sadri of Induction through what gets written to memory, what gets forgotten, and how completed runs and human corrections become new knowledge.