Anthropic explains the components of agent memory and "dreaming"

Anthropic published a post on November 6, "The components of agent memory", on how it thinks about memory for Claude agents and how "dreaming" helps. Anthropic defines agent memory as what Claude learns from experience and keeps between sessions as notes it can read later; weight changes, in-context learning and compaction are outside that definition. It describes dreaming as the agent reflecting on its logs and transcripts between sessions to fix mistakes, write new notes and build new representations of memory.

Findings Anthropic reports from its own evaluations:

• Learning loops. On a database task from Continual Learning Bench (CL-Bench) 1.0, run with Claude Managed Agents, Anthropic says Claude Opus 5 was the first model to close the loop: record a failure, test theories, and turn them into a rule. It says Opus 5.5 and Sonnet 5.5 also closed it.
• Judging what to save. In Pokémon Red, Anthropic says Claude Sonnet 3.7 wrote 413 memory files and earned one badge, while Fable 5 wrote nine files plus a global.md and earned four.
• Steering. Telling Claude only that a memory folder exists was enough for Opus 4.7 and Opus 5. A rule to "record only failures" cut much of their gain from memory. Prompting Sonnet 4.6 to read its memory and test its theories nearly doubled its score.
• Correcting memory. Dreaming checked a wrong note about a Pokémon Center's location against the session transcript and rewrote it, and the run with the corrected note got through Mt. Moon where the uncorrected one did not.
• Wikis. Using about 350,000 files (roughly 420M tokens) from Stainless, which Anthropic says it acquired in May, it had Opus 5 agents build a wiki over 128 sessions. Performance on 25 questions rose from 13% to 48% of graded criteria. Search plus the wiki scored 61% against 53% for search alone and saved about 20,000 tokens per question. On some legal-brief tasks the wiki showed no clear gain.

Anthropic's post on X announcing it had 1,432 points by the evening of November 6.