Holly Herndon and Mat Dryhurst’s Play from Memory (2024) is a series of machine-generated works informed by Gunild Keetman and Carl Orff’s Schulwerk, a set of music education methods and “imitation games” for children developed in the 1920s. Orff and Keetman hoped to teach children to improvise freely from memory, training them with instructions in the form of text prompts, mnemonic games, and graphic symbols. Herndon and Dryhurst created sixteen new machine learning models that produce fantastical orchestral / choral sounds and imagery of imaginary instruments, performances, and lessons. Play from Memory highlights striking similarities between the ways in which Orff and Keetman aspired for children to learn music, and how we train machines to learn music today.