Anthropic Engineering on November 6, 2026 published The components of agent memory, a first-party research essay on how Claude can keep knowledge across sessions. The piece is written by Lance Martin with editorial guidance from Caitlin Love, Tom Chivers, Rebecca Hiscott, Jake Eaton and Blake Wilson, and credits collaborators across Anthropic and outside researchers including work on Claude Plays Pokemon.
The post defines agent memory as knowledge Claude gains from experience and keeps between sessions as readable context, not weight updates, ordinary in-context learning, or compaction summaries that replace older turns. It argues for a split similar to fast hippocampal capture and slower consolidation: agents log work in transcripts, then use idle time for what researchers at Letta and UC Berkeley called sleep-time compute, here framed as dreaming that replays logs to fix mistakes, write notes and build new memory structures.
In-session, Claude can tag salient events or write memory notes as predictions about what future sessions will need. Anthropic reports that newer Claude models got better at judging salience without rigid rules. On Continual Learning Bench (CL-Bench) 1.0 database tasks with Managed Agents, stronger models closed a learning loop by storing tested rules and trap warnings; telling models to record only failures hurt the memory boost. On Pokemon Red, Sonnet 3.7 wrote hundreds of files and progressed slowly, while Fable 5 kept a small folder of plans and rules and earned more badges.
Dreaming is the offline pass. Anthropic shows a Mt. Moon navigation case where an incorrect Pokecenter location note caused repeated failures until dreaming checked the note against the transcript and rewrote it. Dreaming also builds wiki-style linked pages over many sessions: on a Stainless internal corpus of about 350,000 files, scaling dreaming from 2 to 128 sessions raised graded question performance from 13% to 48%, and pairing wiki with search beat search alone on some multi-document questions. On Harvey LAB legal-brief tasks, goal-tuned dreaming instructions that forced per-contract pages with verbatim terms lifted graded contract facts from 72% to 100% in one example.
The essay is careful about limits. Injected top-of-hierarchy memories spend tokens and can poison every session if wrong, so Claude Code auto memory only loads the first 200 lines or 25 KB of MEMORY.md as an index. Dreaming needs compute, can overgrow into maps as large as the source (a Borges image Anthropic uses), and still leaves open when to write memory live versus offline. Anthropic presents the work as design notes and evaluations, not a new consumer product launch.
For agent and memory desks the post matters because a major lab is publishing concrete harness patterns (notes, dreaming, wikis, injected indexes) with named benchmarks and failure modes. Claims about benchmark lifts, Pokemon badge counts and wiki scores are Anthropic's reported results. Independent press was not required under the direct=1 Anthropic Newsroom path; no conflicting first-party correction was found the same day.