Anthropic Says Newer Claude Models Use Memory Far Better, Helped by Offline 'Dreaming'

The findings matter for companies deploying AI agents on long-running work, where an agent that starts each session from scratch repeats mistakes and wastes computing on rediscovering what it already learned. Anthropic defines agent memory as knowledge Claude gains from experience and saves as files it can read later, as distinct from changes to a model's weights or information held within a single conversation.

On a database question task from a benchmark called CL-Bench, in which Claude was told the correct answer after each attempt and could write notes for later sessions, Claude Opus 5.5 scored 0.838 with memory against 0.655 without it, according to the company's figures. Claude Sonnet 4.6 barely improved, moving from 0.330 to 0.364. Anthropic said the older model generated theories about its mistakes but rarely tested them, while Claude Opus 5 was the first of the models to "close the loop" by recording verified rules, how it knew they were true, and warnings about traps.

A run of the video game Pokémon Red showed the same shift in judgment. About 10,000 steps into the game, Claude Sonnet 3.7 had written 413 memory files across 225 folders, much of it duplicating what was already on screen, and earned one badge. Claude Fable 5 had written 10 files and earned four badges, saving only what it had worked out for itself, such as strategy and tested rules.

The company said this argued for telling stronger models what memory is for rather than how to use it. Instructing Opus 5 to record only failures cut its gain from memory by 0.234, it said.

Dreaming, which builds on a "sleep-time compute" idea proposed by researchers at Letta and UC Berkeley, was tested in several settings. In one Pokémon run, a wrong note about a map location sent the agent off a one-way ledge three times; after a dreaming pass checked the note against the transcript and fixed it, the agent reached its destination.

On a corpus of about 350,000 files from the developer tools company Stainless, Claude Opus 5 used repeated sessions to build an internal wiki. Answer accuracy using the wiki alone rose from 13% to 48% of graded criteria over 128 sessions, while combining the wiki with search reached 61%, against 53% for search alone, and saved about 20,000 tokens per question. On 24 merger questions built with legal AI company Harvey, the wiki scored 87% against 73% for search. On a task that needed only a single document, it showed no clear gain, and in one case the wiki grew to 1.77 times the size of the material it was meant to summarise.

The techniques are tied to products Anthropic already sells. The research used Claude Managed Agents, whose memory directory persists files between sessions, and the company points developers to documentation for a dreaming feature and to Claude Code's automatic memory, which loads only the first 200 lines or 25 KB of its index file.

Anthropic said open questions remain, including whether agents should write memories in the moment or in the background, how to keep stored knowledge current, and how often to run dreaming given its computing cost. "Memory systems will evolve as Claude gets more capable, and we still have a lot to learn about designing Claude's memory systems," the company wrote.