Memory-augmented policies for long-horizon tasks

Long tasks fail when policies have no notion of what happened sixty seconds ago. We add a short-term episodic buffer keyed by object identity and a small latent cache for recurring subgoals.

The episodic buffer stores object identities and last-seen poses within a 30-second window. The latent cache compresses recurring subgoal patterns into a fixed-size embedding. Together they lift completion rate from 52% to 88% on our placeholder benchmark.

This post is a filler milestone. Real numbers replace these once we close the first customer trial.