In a post on its engineering blog on 6 November, Anthropic described experiments testing what each component of Claude’s memory adds to an agent’s work. Claude writes notes as it works. When dreaming, it later reviews logs from earlier sessions, corrects its notes and compiles new ones from them.

Strong models benefited from choosing what to remember

In the database task of the Continual Learning Bench 1.0 benchmark, each question started in a new session, but the memory directory persisted. Of the models tested, Opus 5 was the first to record verified rules and how it knew they held.

Strong models should be told what memory is for, but not how to use it, the company said. When Opus 5 and Opus 4.7 were given a rule to record only failures, they lost much of the benefit that memory gave them. With the older Sonnet 4.6, a prompt to read its memory before answering and to test its theories nearly doubled its score.

Dreaming corrects and compiles

In Pokémon Red, Claude recorded a landmark on the wrong side of Mt. Moon and repeatedly ended up back on the previous route. While dreaming, it compared the note with the session log and corrected the error. In a single run, Claude got past the mountain with the corrected note but not with the uncorrected one.

Dreaming also lets Claude compile a wiki: a collection of linked pages that it maintains itself. Opus 5 agents built a wiki over 128 sessions from about 350,000 internal files at Stainless. With the wiki alone, Claude met 13 per cent of the grading criteria for 25 questions after two sessions and 48 per cent after 128. Combining search with the wiki raised the share from 54 per cent to 61 per cent as the wiki grew, compared with 53 per cent for search alone. The combination saved about 20,000 tokens per question.

On M&A questions that required several documents, Claude scored 87 per cent with the wiki and 73 per cent with search alone. In one task where the criteria were often found in a single document, the wiki gave no clear advantage. In another, the wiki grew larger than the documents themselves.

When several agents wrote to the same contract-terms table, they overwrote each other’s work, and 41 of the 57 assessed contract facts survived in the wiki. Claude rewrote the compilation instructions based on its mistakes: each contract got one page written by a single agent, with the terms quoted word for word. All 57 then survived.

Keep loaded memory small

Dreaming can also promote notes to the level that the harness, the program that runs the agent, loads at the start of every session. Anthropic warns that this consumes tokens, that a faulty note can carry into every session and that long instruction files are followed less closely. Claude Code’s automatic memory loads only the first 200 lines or 25 KB of MEMORY.md.

Dreaming requires compute, and checking all memories continuously may not be practical, the company said. Anthropic has published example guides (cookbooks) for setting up dreaming and wikis.