Anthropic explains how Claude’s memory “dreaming” works

Anthropic defines agent memory as knowledge Claude gains through experience and retains between sessions in context it can read later.[1] The company distinguishes that memory from changes to model weights.[1] Anthropic says agents capture much of their work in logs or transcripts, and Claude may tag salient events while a session is underway or write them to a memory store, such as a folder of notes.[1] In Anthropic’s framing, writing down a rule is a prediction that the pattern it describes will matter again.[1]

Anthropic says its Dreams feature lets Claude review past sessions to curate an agent’s memory and surface new insights.[3] A dream reads an existing memory store alongside past session transcripts, then creates a reorganized store that can merge duplicates, replace stale or contradicted entries and surface insights.[3] The input store is not modified, so users can review or discard the output.[3] Anthropic documents Dreams as a research preview, and says it is not eligible for zero data retention or HIPAA.[3]

Anthropic says benchmark questions began in fresh sessions, while Claude could retrieve files from a memory directory that persisted between sessions.[1] It reports that Claude Opus 5 was the first tested model to close the learning loop; Opus 5.5 and Sonnet 5.5 also used memory to record validated rules for future questions.[1] The company’s chart lists mean rewards of 0.655 and 0.838 for Opus 5.5.[1] These results describe the reported benchmark setup, not a general guarantee about performance on other tasks.

Anthropic compared one Pokémon Red run each from Sonnet 3.7 and Fable 5 at about 10,000 steps to examine what the models chose to remember.[1] Sonnet 3.7 created 413 memory files across 225 folders and earned one badge; Anthropic says its detailed record was continually going out of date.[1] Fable 5 wrote nine files in two folders, plus a top-level `global.md`, and earned four badges.[1]