The accelerator in this module computes a network. Learning on the device asks a different question, which is what the network has to carry with it. A 2026 paper trains a Q-learning agent to play Pong from pixels on an ESP32-S3, a chip with 512 kilobytes of on-chip memory, with the game simulated, rendered and learned on the chip itself. Its selling point is that it needs no replay buffer, and a buffer of a million frames would be a gigabyte, so that point is real. But removing the buffer does not leave you with just a network. The appendix lists exactly what the learner keeps between steps: the parameters, an eligibility trace the same shape as the parameters, an optimizer that keeps an element-wise running maximum the same shape again, and statistics for normalising every one of the 2,048 input pixels. It reports the total as about 70 kilobytes and never says how many bytes one number takes. That omission turns out to be a gift. Only one width makes the ledger land on the paper's own figure, and once you know it you can see where the memory goes, which is mostly not the model.