Anthropic measures agent memory: a persistent memory folder lifts Claude Opus 5.5 from 0.655 to 0.838

On November 6, 2026, Anthropic published “The components of agent memory,” an engineering study that lays out a memory hierarchy (raw transcripts, memory notes, representations such as wikis, and injected context) and a “dreaming” process that reorganizes memory offline by reviewing past transcripts. It backs the framework with four internal evaluations.

On a CL-Bench database task answered in fresh sessions, mean reward with a persistent memory directory versus stateless rose from 0.655 to 0.838 for Claude Opus 5.5, 0.644 to 0.833 for Sonnet 5.5 and 0.564 to 0.828 for Opus 5, but only 0.330 to 0.364 for Sonnet 4.6. Prompting Sonnet 4.6 to read memory and test its theories raised it to 0.641. Restricting memory to “record only failures” cut the gain by 0.234 for Opus 5.

In a wiki built from about 350,000 internal files (roughly 420 million tokens), wiki-only accuracy on 25 questions rose from 13 to 48 percent over 2 to 128 sessions, and wiki plus search reached 61 percent versus 53 percent for search alone while saving about 20,000 tokens per question. On legal merger documents, the wiki scored 87 percent versus 73 percent for search on multi-document questions. In Pokemon Red, a dreaming-corrected note let the agent reach Cerulean City where the uncorrected run stalled at Mt. Moon.

Why it matters: memory is the mechanism that lets agents improve across sessions, and the study shows the benefit depends heavily on the model’s own habits of reading and curating memory, not just on providing storage.

What it does not show: many results come from single runs or single corpora and are illustrative; the wiki did not help where single-document search sufficed; some wikis grew larger than their sources (203 KB versus 115 KB in one case); and dreaming and injected memory both cost compute and tokens.

Sources

Last verified November 9, 2026