Anthropic details four-layer agent memory system and dreaming technique

Anthropic's November 6, 2026 research post outlines a four-layer memory architecture for agents and introduces dreaming, an offline process where Claude reviews past sessions to fix mistakes, with benchmark results across several tasks.

Anthropic published research on November 6, 2026, outlining a four-layer architecture for agent memory and introducing a technique called dreaming, in which Claude reviews past session transcripts during idle time to correct mistakes and reorganize what it has learned, as described in Anthropic's post on the components of agent memory.

How the memory system works

A four-layer architecture: transcripts, memory notes, memory representations such as wikis, and injected context.
Dreaming: offline processing where Claude reviews past transcripts during idle time to correct mistakes and reorganize memory, likened to memory consolidation during sleep.
On the CL-Bench database task, mean reward with memory versus baseline rose from 0.655 to 0.838 for Opus 5.5 and from 0.644 to 0.833 for Sonnet 5.5.
In a Stainless-corpus wiki evaluation of 350,000 files and about 420 million tokens, 128 dreaming sessions reached 48% of graded criteria using the wiki alone and 61% with a search tool added, up from 13% after 2 sessions, as the wiki grew from 103 pages and 766 links to 222 pages and 12,893 links.
In a Pokemon-game memory test, Sonnet 3.7 earned 1 badge using 413 memory files, while Opus 5.5 beat the game.
On the Harvey LAB legal-brief benchmark, merger-question accuracy rose from 73% to 87% with a wiki, and a contract-terms task improved from 72% (41 of 57 facts) to 100% (57 of 57 facts) after instruction refinement.

What it means for AI builders

Memory architecture is a core bottleneck for agents that run over long sessions, so these published techniques and benchmark numbers give agent builders a concrete design pattern to borrow, along with a sense of where memory systems currently top out on tasks such as legal document review, database querying, and long-running games.

Sources

The components of agent memory (Anthropic, November 6, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.