Anthropic says the key to agent memory is knowing what to forget, not how much to store
Anthropic argues that long-running agents fail less from a lack of storage than from writing down the wrong things. Because each new session still starts with an unchanged model, Claude keeps knowledge by writing notes to a folder it can reread later. Every note is a bet: the agent must decide, before it knows what comes next, what might matter later.
In a Pokémon Red test, Claude Sonnet 3.7 treated almost everything as important. Across about 10,000 game steps it created 413 files in 225 folders, often one per battle or item. The archive went stale as the game state changed; at times the model only reorganized notes instead of playing, and it finished with a single badge. Claude Fable 5 kept a handful of strategy files, skipped facts already on screen, and recorded plans, strategies, and tested rules, earning four badges. Anthropic compared only one playthrough per model.
Rigid prompts such as “only record mistakes” erased most of the advantage for stronger models; older models did better when told to read their notes before answering. Faulty notes are the larger risk. Dreaming, building on Letta and UC Berkeley work and analogous to sleep, has Claude review earlier transcripts during a break and check notes against them. After wrongly placing a landmark on a mountain, Claude repeated the same trap three times and nearly quit; a Dreaming pass caught the error. Dreaming also built a linked wiki from about 350,000 Stainless documents, working best with search. In a legal test, a wiki scored 87 percent on questions spanning multiple documents versus 73 percent with search alone; simple lookups were roughly even, and one wiki grew larger than the sources.
Notes mid-task distract, Dreaming costs extra compute and arrives late, and wikis need upkeep. Loading too many memories into every session wastes context and spreads bad notes. Anthropic introduced Dreaming in May as a research preview for Claude Managed Agents: it reads an existing store and up to 100 past sessions, cleans duplicates and outdated entries, and writes a new store without touching the original. At launch it supported only Claude Opus 4.7 and Sonnet 4.6. OpenAI also renamed ChatGPT’s June memory overhaul “Dreaming,” maintaining a user profile and correcting outdated facts in the background.
Continue in the app — vote & join in ➔