Anthropic's Claude Managed Agents can now 'dream,' sort of

Source: Ars Technica

Overview

At its Code with Claude developers’ conference, Anthropic introduced “dreaming” for Claude Managed Agents. In this context, dreaming is a process that reviews recent events and identifies specific items worth storing in “memory” to inform future tasks and interactions.

Managed Agents

Managed Agents are a higher‑level alternative to building directly on the Messages API. Anthropic describes them as a “pre‑built, configurable agent harness that runs in managed infrastructure.” They are intended for scenarios where multiple agents collaborate on a task or project over several minutes or hours, rather than a single, short‑lived interaction.

Dreaming Feature

Dreaming is currently available in research preview and limited to Managed Agents on the Claude Platform. It operates as a scheduled process: sessions and memory stores are reviewed, and specific memories are curated. This helps mitigate the limitations of LLM context windows, ensuring that important information isn’t lost during lengthy projects.

On the chat side, many models employ a similar process called compaction, where lengthy conversations are periodically analyzed to remove irrelevant information while retaining what’s essential for the ongoing conversation, project, or task.