Are million-token context windows taking us in the wrong direction?

Modern Large Language Models can reason remarkably well, but they still suffer from one fundamental limitation: they don't retain knowledge between sessions. The common response has been to increase c