Anthropic raised the accuracy of memory-enabled agents from 54% to 61%

Most systems with “memory” actually just dump old conversations into a folder. After several weeks, the agent gains not an advantage but a digital attic: contradictory notes, stale states, and dozens of files that must be reread before every action.

Anthropic proposes a more useful model. The agent should save not everything it has seen, but what will be useful in the next session: verified rules, the current goal, a working plan, and pitfalls that have already cost it mistakes. Then, while idle, it returns to its work logs, checks its notes, and restructures them. The researchers call this process dreaming.

Memory is a forecast, not an archive

When an agent records a rule, it is effectively making a bet: this information will be needed again, and retaining it will be more efficient than retrieving it again. Memory quality therefore depends primarily on the ability to distinguish what matters from what is temporary.

This is clearly demonstrated by an experiment with Pokémon Red. Claude Sonnet 3.7 saved nearly every encounter, battle, and item it found. By around 10,000 steps, it had created 413 files across 225 folders but earned only one badge. Some notes quickly became obsolete, while the agent spent an increasing number of actions maintaining its own archive.

Another model, Fable 5, limited itself to nine files in two folders and one main document. That document contained the current goal, a walkthrough plan, and rules for managing memory. Separate notes described routes and verified gameplay techniques. This agent earned four badges: it had less memory, but every entry helped it make the next decision.

The practical principle is simple. There is no need to save coordinates, a balance, or task status if the agent can obtain them from a reliable tool right now. What should be saved is what the external system does not contain: lessons from mistakes, confirmed patterns, plans, and the reasons behind decisions.

“Dreaming” corrects confidently held errors

The danger of persistent memory is that an incorrect note outlives the session that produced it. On the next run, the agent treats its own note as a source of truth and makes the wrong decision again.

During one Pokémon playthrough, Claude incorrectly recorded that the healing center was on the eastern side of the mountain. As a result, it concluded that it had already cleared the cave, headed in the wrong direction, and returned to the same route several times. After the third repetition, the agent effectively gave up.

During “dreaming,” the model compared the note with the original log, detected the discrepancy, and corrected the location of the object. On the next run, Claude returned to the cave and reached a new town. The important point is not the game story itself but the engineering pattern: the action log remains the primary evidence, while memory is an editable derivative that must be checked periodically.

What this means for production agents

Anthropic tested the same approach on a collection of approximately 350,000 internal documents totaling around 420 million tokens. The agents progressively built an interconnected knowledge base: they created pages, merged duplicates, corrected claims, and verified them against sources. When background processing increased from two to 128 sessions, performance on the question set rose from 13% to 48% of the evaluation criteria.

However, the knowledge base did not replace search. The best result came from combining the two mechanisms: the accumulated knowledge map indicates where to look, while the search tool makes it possible to verify details in the primary source. As the knowledge base developed, the score for this combination rose from 54% to 61%, while search alone scored 53%.

For developers, the takeaway is highly practical. Start with a small memory: one file containing the goal and next steps, a separate section with confirmed rules, and links to the original logs. Add a background task that looks for contradictions, flags unverified conclusions, and removes outdated information. Do not give a powerful model dozens of rigid rules about what to remember: in Anthropic's experiments, requiring it to save only mistakes reduced the value of memory.

The metrics were obtained on Anthropic's own models and small task sets, so the percentages cannot be generalized to every product. But the architectural conclusion appears robust: good agent memory is not an infinite log, but a compact, verifiable, and regularly rewritten working model of the world.