[Late catch] Anthropic on agent memory: letting Claude reread its work logs and fix its notes with "dreaming"

Anthropic wrote up its experiments on giving Claude memory that carries across sessions. Two pieces matter most. First, Claude writes important details to a memory folder as it works. Second, during idle time it rereads past transcripts to correct and merge those notes, a step Anthropic calls dreaming. The experiments ran on Claude Managed Agents with a memory folder that persisted between sessions.

The most common problem teams hit when they add memory to an agent is a wrong note that sticks around and causes the same mistake again. According to the post, dreaming let Claude check notes against the original logs and fix bad ones, and on a set of questions about a company's documents, scaling dreaming from 2 sessions to 128 raised the score from 13% to 48%Vendor claim. Another useful point for design: telling Claude what memory is for worked better than spelling out how to use it.