Anthropic on agent memory: “dreaming” to organize experience, fix mistakes, and build knowledge maps
Anthropic’s Engineering team described how it designs agent memory for Claude: using idle-time “dreaming” to review logs, correct bad notes, and build searchable wikis, with evaluations showing how memory helps on long-horizon tasks.
Anthropic published “The components of agent memory” on its Engineering blog on November 6, 2026 (Singapore time), explaining how the company thinks about and evaluates memory for AI agents.
By “agent memory,” Anthropic means knowledge Claude gains from experience—including user interactions and assigned tasks—that it retains between sessions and can read later. That excludes weight updates, in-context learning within a single conversation, and compaction (summaries that replace older turns so work can continue in a new context window).
The post draws an analogy to how human sleep consolidates memory. Agents already capture work in logs; during idle time they can use “dreaming” to reflect on those logs, fix mistakes, write new notes, build new memory representations, or decide what to preload into future sessions. Anthropic says this builds on sleep-time compute ideas from researchers at Letta and UC Berkeley, and that dreaming has been useful for giving Claude effective memory.
On evaluations, the team used Claude Managed Agents on Continual Learning Bench (CL-Bench) 1.0 and other tasks to see how models form learning loops with a persistent memory directory, and compared what different models judged worth saving in long-horizon settings such as Pokémon Red. Newer models tended to store strategy and validated rules rather than mirroring every on-screen fact.
For wikis built via dreaming—pages Claude writes and links together—the team tested a corpus of about 350,000 internal files (~420M tokens) from Stainless, a company Anthropic acquired in May, with a 25-question set. As dreaming scaled from 2 to 128 sessions, wiki-only scores rose from 13% to 48% of graded criteria. With both search and the wiki, aggregate scores rose from about 54% to 61% as the wiki grew, versus about 53% with search alone, and using both also saved about 20,000 tokens per question. Other tests (including merger Q&A) showed clearer gains when answers needed information across many documents; when criteria lived in a single searchable document, the wiki often added little.
Dreaming can also promote a small set of notes to the top of a memory hierarchy so the harness injects them at the start of every session. Anthropic notes that injected context costs tokens and that a wrong note can follow Claude into every session, so the top layer must stay small. It cites Claude Code auto memory loading only the first 200 lines or 25 KB of MEMORY.md as an index.
This is an engineering research post, not a standalone product-launch announcement. Official post: The components of agent memory