Fixing agent memory

Are we thinking about large language models all wrong? We keep expecting them to somehow be superhuman, yet they regularly reflect all-too-human tendencies. Like when Claude made up passages from old Yorkshire wills that seemed to connect my ancestor to England.