Anthropic has published "The components of agent memory," an engineering post that lays out how Claude agents retain knowledge between sessions and introduces "dreaming," an offline process in which an agent reviews its own logs to fix mistakes, write notes and build longer-lived memory. The post, written by Lance Martin and published on November 6, 2026, compares the idea to how the brain consolidates memories during sleep.
Anthropic defines agent memory as knowledge Claude gains from experience and keeps between sessions as readable context, explicitly excluding weight updates, in-context learning and compaction. It describes a four-layer hierarchy: searchable session transcripts, memory notes Claude writes itself, higher-level representations such as linked wiki pages, and injected context loaded at the start of every session. Dreaming can operate across all four, correcting notes against transcripts and promoting useful notes upward.
The post reports results on Continual Learning Bench (CL-Bench) 1.0. On a database task, mean reward with memory rose from 0.655 to 0.838 for Opus 5.5, from 0.644 to 0.833 for Sonnet 5.5 and from 0.564 to 0.828 for Opus 5. Older models gained less: Sonnet 4.6 moved only from 0.330 to 0.364, reaching 0.641 when memory was paired with refined instructions. Anthropic notes that Sonnet 4.6 read its memory before answering less than half the time, while Sonnet 5 did so nine times out of ten.
On a corpus of roughly 350,000 files (about 420 million tokens), an Opus 5-maintained wiki improved over 128 sessions from 13% to 48% of graded criteria on its own. Wiki plus search reached 61%, against 53% for search alone, while saving about 20,000 tokens per question. In a contracts test, per-contract wiki pages raised graded facts from 41 of 57 to all 57. A Pokémon Red experiment illustrated salience: Sonnet 3.7 wrote 413 memory files and earned one badge, while Opus 5.5 kept a small goal-tracking folder and finished the game.
For developers, the post points to concrete building blocks: the memory tool, a memory directory and dreaming in Claude Managed Agents, and knowledge-wiki and dreaming cookbooks, with a wiki quickstart on GitHub. It also notes a practical limit for Claude Code users: auto memory loads only the first 200 lines or 25 KB of MEMORY.md, so what sits at the top of that file matters.