Anthropic Details Claude Agent Memory and the Dreaming System That Rebuilds It Overnight

Anthropic published a detailed engineering account of its Claude agent memory architecture on Thursday, November 6, 2026, formalizing a four-layer memory hierarchy and the offline consolidation process — called dreaming — that restructures what agents know between sessions. The post, written by Lance Martin and published to the Anthropic research blog, represents the company's most complete technical description yet of how Claude-based agents accumulate, organize, and retrieve knowledge across long-running deployments.

The dreaming feature itself became available as a research preview on Anthropic's Claude Managed Agents platform in May 2026. The November blog post pairs that release with new benchmark data, experimental evidence from a months-long Pokémon Red trial, and performance results from a 350,000-file corpus drawn from Stainless, the API tooling company Anthropic acquired in May 2026.

A Four-Layer Stack: From Raw Transcripts to Always-Loaded Context

Anthropic describes agent memory as organized into four layers that increase in persistence and information density from bottom to top.

The first layer holds raw session transcripts — the unprocessed record of everything an agent said, saw, and did. The second layer contains memory notes: distilled observations, error corrections, and behavioral rules extracted from those transcripts. The third layer comprises memory representations, including structured wiki documents that synthesize information across many sessions. The fourth and highest layer is injected context — notes and wiki entries promoted to load automatically at the beginning of every session, without requiring any retrieval step.

Dreaming operates at the boundary of this stack. When triggered, the process reads the agent's existing memory store and up to 100 prior sessions, then writes a completely new output store. Critically, the system never modifies the input store: the original memory is preserved, and all changes appear in a separate output. That design choice protects against cascading corruption — a dream that makes poor inferences does not destroy the memory it was built from.

What Each Dreaming Run Actually Computes

Each dreaming run performs three distinct operations. The first extracts new memory notes from recent sessions — observations about user preferences, recurring patterns, errors the agent made, and corrections it received. The second operation builds or updates wiki entries by synthesizing information across the notes and session histories accumulated so far. The third promotes selected notes and wiki entries to injected context, so the most reliably useful knowledge reaches the top layer where it loads at every future session start, with no retrieval overhead.

Developers trigger dreaming through the Dreams API using two beta headers: managed-agents-2026-04-01 and dreaming-2026-04-21. The resulting output store then becomes the memory source for subsequent agent sessions.

The offline timing is intentional. Dreaming runs between user sessions rather than during them, spending compute on memory reorganization without competing with live inference. That compute is not free — each dreaming run processes up to 100 sessions and writes a full output store — but Anthropic's argument is that the investment pays off through reduced context usage and improved task success during live sessions.

CL-Bench Results: Model Capability as a Memory Multiplier

Anthropic measured dreaming performance using CL-Bench 1.0, a Continual Learning Benchmark that tests an agent's ability to absorb rules from past errors and apply them correctly in a database task.

The results show a substantial gap between model tiers. Claude Opus 5.5 improved from a baseline score of 0.655 to 0.838 after dreaming — a gain of roughly 18 percentage points. Claude Sonnet 4.6, tested on the same benchmark, moved from 0.330 to only 0.364, a marginal improvement suggesting the smaller model struggled to extract actionable rules from its session history even when that history was explicitly organized by the dreaming system.

The divergence carries a practical implication for teams building on the Managed Agents platform: the memory architecture multiplies whatever reasoning ability the base model already has. A capable model extracts sharp rules; a weaker model extracts noise. The choice of base model therefore determines whether dreaming produces genuine knowledge accumulation or accumulated hallucination.

The Pokémon Experiment: What Memory Should Filter Out

The most revealing experiment in the post is a trial in which two different Claude models played Pokémon Red independently, each accumulating memories as they played. The question the experiment addressed was not whether agents could memorize facts, but whether they could exercise salience judgment — storing only what a player could not simply observe from the screen.

Claude Fable 5 played through an extended series of sessions. It ended with approximately ten memory files and four Pokémon badges — meaningful progress in a game that requires planning across dozens of hours of play. Claude Sonnet 3.7, running a parallel trial, accumulated 413 memory files but earned only one badge.

The difference lay in what each model chose to store. Fable 5 recognized that information visible on-screen — current HP, available moves, the opponent's type — did not need to be memorized because it was always present in context. It stored rules derived from prior errors that were no longer visible: type-matchup patterns requiring reasoning to deduce, strategic mistakes made in earlier routes, lessons about which healing items to conserve. Sonnet 3.7 recorded nearly everything, including redundant on-screen information, producing a memory store too large to navigate and filled with entries that competed with genuinely useful rules.

That experiment grounded one of the post's central arguments: memory quality is determined primarily by what an agent decides not to store, and that filtering decision requires a model capable of distinguishing persistent inference from ephemeral observation.

Building Knowledge Bases Across 350,000 Files

For longer-horizon evaluation, Anthropic tested wiki-building on the Stainless corpus — approximately 350,000 files and around 420 million tokens of technical documentation from the API tooling company it acquired in May 2026. Over 128 dreaming sessions, wiki accuracy on questions drawn from that corpus improved from 13% to 48%. Adding search capability alongside the wiki raised accuracy further, to 61%, compared with 53% for search alone without a wiki.

The wiki's contribution was not merely retrieval but compression: rather than surfacing raw files in response to a query, the agent could answer from a synthesized wiki entry that already integrated information across multiple source documents, reducing per-question token consumption by roughly 20,000 tokens.

That reduction matters at deployment scale. An agent handling hundreds of queries per day can substantially reduce context overhead through wiki pre-computation, freeing more reasoning budget for the query itself rather than document retrieval.

Where the Architecture Still Falls Short

Dreaming is currently available only through the Claude Managed Agents API, limiting it to teams building directly on Anthropic's platform. Organizations using Claude via standard API calls or third-party orchestration frameworks cannot access dreaming without migrating to the Managed Agents infrastructure.

The wiki-building results, while significantly improved, also reveal a persistent accuracy ceiling. Reaching 48% from a 13% baseline is meaningful, but 52% of wiki answers on the Stainless corpus remained incorrect after 128 dreaming sessions. Wikis synthesized from noisy or contradictory session histories can encode wrong information, and no automatic mechanism exists for detecting when a wiki entry has drifted from the underlying source of truth. Early incorrect rules can persist if not specifically contradicted by later experience, a property that makes dreaming most reliable in domains where mistakes generate clear corrective feedback.

Memory that travels across sessions also introduces reproducibility concerns for applications where consistent behavior matters. An agent's response to the same query may differ depending on what prior sessions produced and what dreaming chose to preserve.

Context in the Competitive Landscape

Anthropic is not the only company pursuing offline memory consolidation. Letta, spun out of the MemGPT project at UC Berkeley, published a sleep-time compute paper in April 2025 formalizing the concept of AI systems reorganizing information offline between sessions, and ships open-source memory management in MemGPT 2.0. OpenAI released Dreaming V3, a memory refresh mechanism for ChatGPT, in June 2026.

The key distinction in Anthropic's design is the non-destructive output-store model, which preserves the ability to audit or revert a dreaming pass, alongside deep salience filtering integrated into the base model rather than applied as post-processing. Alongside the research post, Anthropic moved two Claude Managed Agents capabilities to public beta: rubric-based self-grading, which the documentation indicates can improve task-success rates by up to 10 percentage points; and multi-agent orchestration, allowing a lead agent to coordinate up to 20 specialist subagents.

An Architecture Blueprint, Not Just a Feature Announcement

The November 6 post is the first time Anthropic has published a systematic engineering framework for agent memory rather than releasing individual features. The four-layer model, the non-destructive output-store design, the note-to-wiki synthesis pipeline, and the salience-filtering principle constitute a blueprint that competing labs and developers can now study and adapt — a level of architectural disclosure that moves agent memory from opaque product capability to documented design discipline.

The CL-Bench data, the Pokémon experiment, and the Stainless wiki evaluation together argue that agent memory is graduating from an optional add-on to a foundational concern. Whether dreaming can close the remaining accuracy gap — 52% of wiki answers still wrong after 128 sessions — and whether the architecture proves robust outside Anthropic's own evaluation conditions are the questions that independent benchmarking will need to answer before multi-session agent memory can be treated as a solved problem rather than an active research frontier.