Anthropic’s Claude Dreams: Agents autonomously organize memories between tasks, eliminating duplicates and contradictions

Anthropic officially unveiled a new feature for Claude Managed Agents called “Dreams” at the Code with Claude developer conference in San Francisco on May 6, entering the Research Preview stage. Anthropic’s official documentation explains that Dreams allows Claude Agents to autonomously organize memories between tasks, eliminate duplicate information, update outdated entries, and uncover new insights. The goal is to ensure that as the agent is used over time, its performance does not degrade due to messy memories. This feature represents Anthropic’s concrete productization in the direction of “Agent self-improvement,” and, together with its May 6 same-day announcement of a compute-power collaboration with SpaceX, it accounts for two major developments unveiled in one day at the Code with Claude conference.

Dreams tackles the problem: memory accumulation leading to noise

Anthropic’s documentation states that when Managed Agents execute tasks, they write the context into a memory store. However, these writes are “local and incremental”—after multiple sessions accumulate, the memory store may develop:

Duplicate entries—identical information written multiple times across different sessions

Contradictions—early writes and later writes are inconsistent, making it impossible to determine which is true

Outdated entries—early user preferences get overwritten by new preferences, but old data is not cleared

Dreams enables Claude to read the existing memory store and up to 100 session records, then generate a brand-new, reorganized memory store—merging duplicates, replacing outdated entries with the latest values, and extracting new patterns. The input memory store is not modified. Users can review the output and decide whether to adopt or discard it.

Asynchronous pipeline: execution takes minutes to tens of minutes, and can be canceled mid-run

Dreams is an asynchronous task; the execution flow is:

The user submits the existing memory store + up to 100 session records, and selects a model (claude-opus-4-7 or claude-sonnet-4-6)

Anthropic’s backend starts the dreaming pipeline; the status moves from pending → running → completed

Execution takes minutes to tens of minutes, depending on the input size

Users can stream in real time to observe what the pipeline is reading and writing

It can be canceled while running; the output store retains the content already produced for inspection

API access is enabled via two beta headers: managed-agents-2026-04-01 and dreaming-2026-04-21. Fees are calculated according to the model token standard. Anthropic sets a 4,096-character limit for instructions, with a maximum of 100 session inputs per request.

What to watch next: research preview access applications and enterprise adoption

Dreams is currently in the research preview stage. Specific trackable events include:

Research preview access applications—must be submitted via the Claude Managed Agents form provided by Anthropic

Model expansion support—currently only claude-opus-4-7 and claude-sonnet-4-6 are supported; whether it will expand to other models later is unknown

Official launch timeline—no information has been released on the timing from research preview to GA

For Claude users and Anthropic’s business strategy, Dreams is a concrete implementation of “the agent gets better the longer you use it.” Previously, Claude Agents were limited by within-session context, and long-term memory organization across sessions required developers to handle manually; Dreams abstracts that layer and turns it into an out-of-the-box capability of the managed platform. Whether this feature can become a differentiating selling point for Claude versus competitors like OpenAI and Google, and whether enterprises choose to adopt it, are worth monitoring.

This article Anthropic rolls out Claude Dreams: Agent autonomously organizes memory between tasks, eliminates duplicates and contradictions first appeared on Chain News ABMedia.