The components of agent memory

Lance Martin borrows a model from human memory: experiences are captured as they happen, and an offline replay later consolidates some of them into lasting knowledge. For an agent, the first part is memory notes Claude writes during a session, which the article treats as predictions about what a future session will need. The second is what Anthropic calls dreaming, reflecting on logs to fix mistakes or produce new notes and representations. The article walks through what dreaming can do: correct a wrong note against the session transcript, build a linked wiki out of loose notes, promote the most important notes into context loaded at the start of every session, focus memory on a specific goal, and fill in gaps a session leaves behind. Among the findings, stronger models judge what is worth remembering better than fixed rules do, and on a document question set Claude scored 53% with search alone and rose from 54% to 61% as it built and used a wiki alongside search. It closes with open questions about when to write memory, how to keep it current and what dreaming costs. For us the useful framing is that a memory note is a bet on the future, which makes it worth reviewing and correcting rather than only appending to.