When a robot learns a new behavior, what keeps it from forgetting the ones it already knows? Prior work, especially on VLAs, shows that rehearsing past experiences can enable strong continual learning. But why does rehearsal work so well, and when does it fail? We find that a small but ✨special subset✨ of rehearsed data largely determines whether old behaviors are retained or forgotten. We call these examples Memory Anchors. ⚓ Withholding just the top 10% of Memory Anchors from rehearsal increases forgetting by up to 4.5x. Increasing their presence, meanwhile, reduces task forgetting and enables a real robot to learn challenging task sequences. Website: robot-adaptation.github.io/M… Paper: arxiv.org/abs/2608.26545 Curious? Read on! 🧵👇 (1/9)