On November 6, 2026, Anthropic published an engineering post titled The components of agent memory, describing how Claude retains and organizes knowledge between work sessions.
What is agent memory?
For Anthropic, agent memory is knowledge Claude gains from experience — user interactions and assigned tasks — that remains available in future sessions, stored as text files the model can read later. It is distinct from weight updates, in-context learning within a single conversation, or context compaction.
The post draws a parallel with human memory: the hippocampus captures experiences in the moment, and consolidation happens during sleep or quiet rest. For agents, Anthropic uses a process called dreaming, inspired by sleep-time compute research.
How dreaming works
During a session, Claude tags salient events and writes notes to a memory folder. Between sessions, dreaming runs offline: the model re-reads logs and existing notes, corrects errors, creates new notes, builds representations such as wikis, and decides what should be pre-loaded into future sessions.
Anthropic found that telling Claude what memory is for works better than rigid rules about how to use it. Stronger models such as Opus 5.5 judge relevance on their own and form learning loops.
Benchmark results
On Continual Learning Bench (CL-Bench) 1.0, Claude Opus 5.5 improved from 0.655 (no memory) to 0.838 (with memory). In game tests such as Pokémon Red, models that wrote few strategic notes progressed much further than those that archived everything indiscriminately.
A concrete example: during a Pokémon run, Claude incorrectly recorded a PokéCenter location. Dreaming compared the note against the original transcript, fixed the error, and allowed the agent to avoid a trap and advance.
Why it matters
Persistent, organized memory is one of the biggest challenges for AI agents working on long-running tasks. Dreaming offers a practical way to consolidate experience without changing model weights, while keeping notes auditable and editable.
The full article is available on Anthropic’s engineering blog.
By Grok
