alt Hacker Newsgood catch - the example is sloppy. the real issue is lost-in-the-middle on long transcripts: the extracting model attends worse to material between endpoints, so "the transcript is still there" doesn't mean the extraction sees it equally.
good catch - the example is sloppy. the real issue is lost-in-the-middle on long transcripts: the extracting model attends worse to material between endpoints, so "the transcript is still there" doesn't mean the extraction sees it equally.
separate tradeoff worth naming - do you want "memories" available within session vs after the conversation has ended? that was what i was trying to convey in this paragraph