Claude Managed Agents are 'Dreaming'

Ronni Holmvig Strøm · 2026-05-22

On May 6, Anthropic introduced "dreaming" to Claude Managed Agents: a scheduled offline process that reviews past sessions, merges duplicate memories, replaces stale entries, and surfaces patterns a single agent cannot see from inside its own context. Harvey reports a roughly 6x improvement in completion rates on long-form legal drafting with dreaming enabled. Discount that number for vendor optimism and preview-stage selection. The choice of name is what we are not discounting.

The Metaphor Is Architectural

Anthropic could have called this compaction, reflection, or memory hygiene. They called it dreaming. For roughly thirty years, the dominant account of sleep has been that it does memory consolidation: the hippocampus encodes the day's experience at high speed and replays compressed traces during slow-wave sleep, and the cortex slowly integrates them into long-term structure.

The shape is the same. An agent accumulates session-level memory at the speed of work. A separate, scheduled process runs when the agent is idle, looks across the accumulated traces, and rewrites the memory store. The fast loop encodes. The slow loop consolidates. The field has now built that two-loop shape twice, in different substrates.

This is not a claim that Claude dreams. The metaphor is loose at the surface and tight underneath. What it names is that offline processing of stored experience does work no amount of in-context cleverness can do.

Longer Context Was Not Enough

For two years the scaling story on the inference side has been longer context: more tokens in the window, better attention over the haystack. Dreaming is the field admitting, in product form, that this is not the whole answer. A million-token window does not solve the problem of which tokens to keep across a million sessions.

The Ars Technica writeup draws the contrast explicitly against compaction, which trims within a session. Dreaming runs between sessions, and across agents in a multi-agent system. That cross-session posture is what makes it interesting. It is the first widely deployed feature of its kind in a major commercial managed-agent platform.

What Gets Borrowed Next

Consolidation is one move. The neighboring functions in the hippocampal-cortical dialogue are not exotic, and the most obvious next borrowing is adaptive forgetting. The literature is clear that biological memory does not merely consolidate, it actively prunes. An agent that dreams without forgetting accumulates a tidier and tidier pile of everything that ever happened to it.

What Anthropic shipped on May 6 is one feature in one product. What it signals is that agent design and cognitive architecture are no longer two separate conversations.

As a footnote, it is worth mentioning that we have been running a similar setup for our AI entities in Privateers for some months, which we choose to call Dream Engine, with great results. We will elaborate on that in a future long form article.