A new framework called AHC enables continual object detection on microcontrollers with under 100KB of memory. It uses MAML-based compression to adapt to new tasks in five inner-loop steps. By applying scale-aware ratios to FPN redundancy patterns, the system reduces catastrophic forgetting. This allows complex vision tasks to run on severely constrained hardware.