Anthropic details Claude agent memory, salience, and dreaming

Anthropic has published its thinking on agent memory for Claude, describing how memory notes, salience judgment, and dreaming work together across sessions. It covers evaluations on database tasks, Pokémon Red, and document wikis.

Memory and dreaming

Anthropic defines agent memory as knowledge Claude gains from experience, including interactions with users and assigned tasks, that persists between sessions as context it can read later. It excludes changes to model weights, in-context learning within one conversation, and compaction. The company compares this to human memory: the hippocampus captures experiences quickly, while much consolidation into long-term knowledge happens during sleep and quiet rest through replay. For agents, logs and transcripts capture work, and researchers at Letta and UC Berkeley proposed sleep-time compute, using idle time between interactions to reorganize context. Anthropic says the same idea can help agents organize memory by reflecting on logs, often called dreaming.

Salience and learning loops

Claude can tag salient events in a session log as they happen or write them to a memory store such as a folder of memory notes. Memory can also predict future needs, such as a verified fact or rule. On a Continual Learning Bench 1.0 database task, models differed: Sonnet 4.6 generated untested theories, Opus 4.7 consolidated notes and marked checked facts, Sonnet 5 marked confirmed answers and corrections, and Opus 5 closed the learning loop with rules, earlier answers, and warnings about traps. Opus 5.5 and Sonnet 5.5 also closed the loop and scored about the same as Opus 5 with memory, though their baseline without memory was slightly higher.

In Pokémon Red, Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Fable 5 wrote nine files in two folders plus global.md and earned four badges. Sonnet 3.7's indiscriminate records became outdated and consumed time; Fable 5 stored strategy, plans, and tested rules, with global.md acting as a dashboard. Opus 5.5 used a small folder of notes tracking current goals and next steps, rewrote them as it played, and beat the game.

Correcting and building memory

Dreaming can test notes against original transcripts. In Pokémon, an incorrect note said the PokéCenter was on the east side of Mt. Moon; with the uncorrected note Claude dropped back onto an earlier route three times and never reached Cerulean City, while dreaming corrected the note and Claude reached Cerulean City. Dreaming can also build wikis, collections of linked pages. Over 128 sessions on about 350,000 Stainless files and roughly 420M tokens, performance on 25 questions rose from 13% to 48% of graded criteria. Using search and the wiki, aggregate scores rose from 54% to 61% as the wiki grew, compared with 53% for search alone, and the combination saved about 20,000 tokens per question.

On merger questions, Claude scored 87% when consulting the wiki versus 73% with search alone, especially when information spanned several documents. Dreaming can also close gaps: on one task it found nine links to a missing page and repointed them, and added three new pages including an index.

Promotion and goal-focused memory

Dreaming can promote notes into the top of the memory hierarchy, where the harness can load them into every session. This injected layer must stay small because injected memories use tokens and an incorrect note can spread; Claude Code loads only the first 200 lines or 25 KB of MEMORY.md as an index. Dreaming can also shape memory toward a goal. In a contract task, a shared table lost facts because agents overwrote one another; after revised instructions gave each contract its own page with word-for-word terms, graded brief criteria rose from 72% to 100%.

Sources: anthropic.com