The article outlines agent memory as knowledge Claude retains between sessions, using memory notes and a process called dreaming to reorganize or correct stored information. It reports CL-Bench and Pokémon Red evaluations comparing how different Claude models saved, tested and revised their notes.