Anthropic research breaks agent memory into notes, salience judgments and idle-time "dreaming"

Anthropic published a research post on 6 November that sets out how it thinks about memory for AI agents. The company defines agent memory narrowly: knowledge Claude picks up from experience and keeps in readable form from one session to the next. Changes to model weights, in-context learning and the compaction of long conversations are excluded.

Three building blocks

The first component is the memory note. During a task, Claude can flag events or write entries to a memory store, and the post treats each note as a bet that the information will matter again. The second is salience: deciding what to keep before knowing what comes next. Anthropic reports that stronger models made better choices when left free to decide than when given fixed rules. The third is what the company calls "dreaming", an offline pass, likened to memory replay during sleep, that re-reads transcripts and logs to fix mistakes, write new notes or build structured references such as wikis.

To measure the effect, Anthropic introduced CL-Bench 1.0. In its database task, each question ran in a fresh Claude Managed Agents session, and a persistent memory directory was the only thing carried over. With that directory, Opus 5.5 rose from 0.655 to 0.838 in mean reward and Sonnet 5.5 from 0.644 to 0.833. Older Sonnet 4.6 barely moved (0.330 to 0.364); the post notes it consulted its memory before answering less than half the time, and refined instructions brought it to 0.641.

Examples and stated limits

Other experiments used Pokémon Red. One early model wrote 413 files and spent effort reorganising them, while a later one kept 10 files and earned more badges. In another run, dreaming corrected a wrong note about a map location that had trapped the agent in a loop. On a corpus of about 350,000 files from Stainless, a company Anthropic acquired, a wiki built over 128 sessions raised Opus 5's score on 25 questions from 13% to 48% of graded criteria, according to the post. Interesting Engineering, which covered the research, also highlights these game and document tests.

Anthropic is candid about the weak points. Consolidation consumes compute, recording notes mid-task can distract from the work, incorrect or outdated notes may be trusted in later sessions, and in one test the wiki ended up larger than the documents it summarised. The post does not present dreaming as a generally available product feature; it points developers to cookbooks and existing tools such as the Managed Agents memory directory.