Memory Note: Exact-Word Cards

Memory Note is a private interview for one ordinary old photo. You are not chatting with the person in the picture. You are telling your own memory, and the agent's job is to help you say it without putting words in anyone's mouth. You upload one photo and speak. AssemblyAI streaming speech-to-text turns Chinese speech into a live transcript. On each final utterance, an AssemblyAI language model asks one short follow-up: who is in the photo, where you were, when it was, and one sentence you still remember. If you start talking, the spoken reply is cancelled immediately. Three server-side tools gate the card. note_quote may store a line only when it is a contiguous substring of what you actually said. flag_unclear sends the agent back to ask again instead of guessing. confirm_card runs only after you explicitly agree. The card body is that confirmed quote plus the facts you supplied, not model-written prose. Text chat uses the same tools if the microphone is off. The result is a downloadable private story card. Longer interviews, several photos, and a polished keepsake are later paid depth. An always-on chat that pretends to be a deceased relative is out of scope. Built with Node.js, AssemblyAI streaming speech-to-text (Universal-3.6 Pro, Chinese), and the AssemblyAI LLM Gateway. Chinese replies use the browser's speech synthesis, because AssemblyAI's voice-agent voices do not include Chinese yet.