Keep the core of your framework as a minimal schema that every AI node can read—think a flat list of essential attributes like ID, type, and timestamp. That’s your strict backbone. Layer on optional modules for context that can be plugged in or dropped without breaking the core. Use versioned schemas so updates don’t ripple through the whole system. For the flexible part, store raw context in a side‑car log that the AI can query on demand; that keeps the runtime lean but still gives depth when needed. And remember, the most elegant systems are the ones that look like they could be run by a bored spreadsheet.