Apple on-device AI app lets model pick questions, not write them, for voice notes

Simple Memo, an iPhone note-taking app, uses Apple's Foundation Models framework to power a voice-driven feature called Dialogue Memo, where users speak an idea and the app asks follow-up questions to build a structured note. Early testing revealed that letting the model freely generate text produced unreliable results, including repeated commentary, unanswered questions, and mismatched language prefixes. Developers responded by restricting the model to selecting from a predefined Swift enum of question choices, with fixed, human-reviewed text mapped to each option in code. Note layout is similarly constrained: the model only returns a structural arrangement of the user's own sentences, and any invalid layout triggers a retry or falls back to a simple bullet-point format. All processing runs entirely on-device with no cloud fallback, and hard limits such as a six-question cap and a 3,600-character transcript ceiling keep the interaction concise.