Sensors Are the Real Context Window

Language systems reason over a curated context. Physical systems reason over an unbounded, noisy, continuously arriving stream that must be compressed before it can be reasoned about at all.

That compression step — what to keep, at what resolution, for how long — is where most embodied system performance is decided. A perfect planner over a badly compressed state is worse than a mediocre planner over a well-chosen one.

We build explicit memory tiers: a high-rate short-horizon buffer for control, a mid-rate scene graph for task reasoning, and a low-rate episodic store for retrieval. Each tier has its own eviction policy tuned to the task, not to a generic recency heuristic.

This is unglamorous infrastructure work, and it is consistently the highest-leverage change we make when a physical AI pilot is underperforming.