Does it have to know the information for all frames, or just the key frames? If it’s the later, then you could encode so the initial table, has X key frames spread out over a few hours, then force a frame to exist in those locations when encoding the live stream.
