Anthropic "Dreaming": Claude Agents Now Self-Improve Between Sessions

Anthropic released "Dreaming" for Claude Managed Agents on May 6, 2026 — a background process that runs between agent sessions to review past work, prune stale memories, merge duplicate information, and surface patterns that no single session could identify. The system does not change model weights. It curates the agent's persistent memory and context so the next session starts with a cleaner, more accurate knowledge base than the previous one left behind. Harvey, the legal AI platform, saw task completion rates increase roughly 6x after implementing Dreaming. Wisedocs, a medical document review company, cut document review time by 50%.

Those numbers are large enough to require a look at what is actually happening under the hood before accepting them at face value.

What Dreaming Actually Does

Dreaming is a scheduled background process that fires between sessions — when an agent is not actively working on a user task. It operates on the agent's persistent memory store and prior session logs.

The process has four stages:

Review: The system reads the agent's accumulated session records and memory entries, looking at the full history rather than just the most recent session.

Merge: Duplicate information is identified and consolidated. An agent that has been told the same fact by different users in different sessions (a company name, a style guide rule, a data source) will have redundant entries; Dreaming identifies and merges them into a single authoritative record.

Prune: Outdated entries are removed. If a previous session recorded a deadline that has passed, a price that has changed, or a workflow that has been superseded, Dreaming removes or flags those entries rather than leaving them as potential sources of incorrect responses in future sessions.

Pattern extraction: The system identifies patterns across sessions that no single session would see. Recurring mistakes — the same type of error appearing across multiple sessions with different users. Team preferences — choices that multiple agents working in the same environment consistently make. Workflow convergence — different agents independently discovering the same efficient approach to a task type.

The output is a curated memory set that the agent begins its next session with. The underlying model — Claude Sonnet, Haiku, or whichever managed agent variant — is unchanged. What changes is what the model knows about its specific deployment context when it starts.

The Harvey Result: 6x Task Completion

Harvey is an AI platform built for legal workflows — document review, contract analysis, research, drafting. The legal domain has specific characteristics that make Dreaming's memory curation particularly valuable.

Legal work is procedurally dense. The same factual patterns recur across matters (jurisdiction rules, filing deadlines, standard clause structures). The same preferences recur across clients and teams (preferred citation formats, drafting styles, entity naming conventions). Without Dreaming, Harvey agents accumulated this context gradually within each session, or relied on users to re-establish it. With Dreaming, the agent begins each session with the accumulated preferences and patterns from all prior sessions that Dreaming has curated.

Harvey describes the 6x task completion rate increase as driven by agents requiring dramatically fewer clarification rounds and fewer error corrections per task — the agent already knows, from Dreaming's pattern extraction, what preferences apply and what mistakes to avoid. The 6x figure is task completion rate (tasks completed to user satisfaction per session), not raw speed.

The Wisedocs Result: 50% Faster Document Review

Wisedocs processes medical documents — clinical records, insurance claims, prior authorisation requests — using AI to extract structured information from unstructured documents.

Medical document review has its own repetition pattern: the same document types, the same field extraction requirements, the same edge cases (handwritten annotations, non-standard abbreviations, missing data). Dreaming's ability to surface recurring patterns across a large volume of similar documents means agents learn which document structures are most common, which extraction errors recur, and which disambiguation rules apply — and apply that knowledge from the first page of each new document rather than rediscovering it partway through.

Wisedocs reports a 50% reduction in document review time. The practical mechanism: agents that previously spent significant time rechecking and correcting extraction errors now make those errors less frequently, and errors that do occur are identified faster because the agent recognises the pattern.

What This Is and Is Not

Dreaming is a memory management system, not a new kind of learning. The distinction matters for understanding what it can and cannot do.

What it is: A scheduled process that curates an agent's persistent notes and context. The model weights remain identical between sessions. The model's general capabilities do not change. What changes is the quality and relevance of the contextual information the model has access to at the start of each session.

What it is not: Online learning, fine-tuning, or reinforcement learning. Dreaming does not update the underlying model based on experience. It does not give the model new capabilities it did not have before. An agent running on Claude Sonnet with Dreaming is still running on Claude Sonnet — it just has better notes.

The distinction matters practically: Dreaming does not raise AI safety concerns about models developing unexpected capabilities through deployment. It raises different concerns (about the quality and accuracy of accumulated memory, about privacy implications of session data being retained and analysed across users) that are more analogous to database management than model training.

Availability and How to Access It

Dreaming is available in research preview for Claude Managed Agents as of May 6, 2026. It is not available for standard API users using the Messages API directly. It is specifically a Managed Agents feature — meaning it requires using Anthropic's agent orchestration layer rather than building custom agent loops on top of the raw API.

The research preview designation means Anthropic is collecting feedback from enterprise customers on memory quality, pruning accuracy, and unexpected behaviour from the curation process. Harvey and Wisedocs are among the first production deployments Anthropic has cited publicly.

For developers building agent systems: Dreaming is the clearest signal yet that Anthropic's product roadmap for enterprise AI is centred on Managed Agents as the primary deployment surface, not raw API access. The features that matter for production AI — persistent memory, cross-session learning, autonomous improvement — are being built into the managed layer.

Why This Matters Beyond the Benchmark Numbers

The Harvey 6x and Wisedocs 50% numbers are enterprise-specific. Your deployment context will produce different results depending on how repetitive your agent's task domain is and how much accumulated session data is available to extract patterns from.

The more important signal is architectural. Dreaming represents Anthropic's answer to one of the core limitations of stateless language model APIs: every session starts from scratch. Enterprises building AI agents for real work have been solving this problem with custom memory systems, vector databases, and context injection pipelines. Dreaming is Anthropic offering that infrastructure natively within the Managed Agents product.

The competitive context: OpenAI's Deployment Company (launched May 11) puts 150 forward deployed engineers inside enterprise clients. Anthropic's response appears to be making the underlying agent infrastructure smarter so fewer human implementation hours are required. The two approaches — more human implementation vs. more autonomous agent improvement — will compete on enterprise outcomes over the next 12-18 months.

Key Takeaways

• Dreaming released May 6, 2026: Available in research preview for Claude Managed Agents; background process that runs between sessions, not a model update
• What it does: Reviews past sessions, merges duplicate memory, prunes outdated entries, extracts patterns across sessions that no single session sees
• Harvey (legal AI): 6x task completion rate increase — fewer clarification rounds, fewer error corrections per task
• Wisedocs (medical docs): 50% reduction in document review time — agents apply accumulated extraction patterns from first page
• What it is not: Online learning, fine-tuning, or weight updates; the model is unchanged; it is memory curation, not new capabilities
• Architectural signal: Persistent, self-improving memory is a Managed Agents feature, not a raw API feature — Anthropic is building the intelligent infrastructure layer into its enterprise product

For Anthropic's infrastructure expansion that provides the compute for these agent workloads, read Anthropic Leases SpaceX Colossus 1: 220K GPUs, Claude Rate Limits Doubled. For the banking AI agents Anthropic launched with Goldman Sachs and Blackstone, read Anthropic + Goldman Sachs + Blackstone $1.5B JV: Banking AI Agents Are Live.