Satellites learn new skills without forgetting old ones — a problem that matters as orbit fills with computing

Researchers built a system that lets satellites in low-Earth orbit learn to recognize new types of objects over time without losing the ability to recognize old ones. This matters because satellites have almost no memory or power, but need to keep learning as new missions emerge.

Satellites in low-Earth orbit are beginning to do their own computing instead of just sending data home. The problem: when a satellite learns to identify a new class of objects (say, ships instead of just buildings), it tends to forget how to do the old task — a limitation called catastrophic forgetting that gets worse under the memory and power constraints of space. This paper describes a system that splits the forgetting problem into three separate parts and solves each one, which means satellites could accumulate skills over their lifetime instead of needing a complete retrain. The constraint is real: orbital constellations like Starlink have hundreds of units, and retraining everything on the ground defeats the purpose of on-board computing.