This seems very close to Langchain's "summary" memory functionality (which seems to have existed since March 2023?) - granted I haven't read the paper just yet