Agent memory: how Claude repairs wrong notes with dreaming

Claude, an AI agent, sometimes carries incorrect information across multiple sessions. Anthropic, the company behind Claude, published an article on November 6 titled "The components of agent memory" that details how this issue was addressed. The article uses a Pokémon Red example to illustrate the problem. In the game, Claude initially believed that a Pokémon Center was on the east side of Mount Moon when it was actually on the west side.

This misinterpretation led to a wrong path, causing Claude to walk onto a one-way ledge and return to the previous location, failing to reach its destination. To address this issue, Anthropic uses "dreaming," which refers to additional sessions that review transcripts and reorganize persistent context. During the dream session, Claude compared the incorrect note against the original session transcript, identified the error, and corrected the memory.

This correction allowed Claude to successfully navigate from Mount Moon to Cerulean City. Anthropic defines agent memory as knowledge retained between sessions, which can be stored in various ways, such as folders of files or context the agent can read later. The key point is that the persistence mechanism has no inherent knowledge about the truth of the information stored.

The article highlights the importance of distinguishing between current observable state and knowledge that needs to survive, as well as the potential benefits of organizing information more efficiently.