If you follow AI news, you may have heard about the “Code with Claude” conference Anthropic held in San Francisco on May 6, 2026. Among the day’s announcements, a feature called Dreaming stood out: it lets an AI agent review its own mistakes between work sessions and adjust its memory automatically, with no human involvement.

The idea is simple, but the results are already striking. Without changing a single line of code, and without retraining any model, companies using Claude agents have seen performance improve session over session. Harvey, a company building AI tools for law firms, reported roughly a 6x increase in its task completion rate after turning Dreaming on. That number is worth pausing on.

1. The problem Claude Dreaming solves

Until now, the biggest weakness of AI agents in a professional setting was the lack of persistent memory from one session to the next. An agent could complete a task correctly one day, then repeat the same mistake the next, simply because it had no record of what happened before.

Take a concrete example: an agent that helps draft legal contracts. If that kind of document requires a specific format, references to certain clauses, or a particular writing style, the agent had to relearn those details at every new session. The result: time lost fixing repetitive errors, more back and forth with users to clarify parameters, and lower productivity.

Dreaming tackles this problem head on. It’s a scheduled process that runs in the background while the agent is idle, a bit like a night’s sleep that consolidates the day’s learning. This process reviews previous sessions, identifies recurring patterns (common mistakes, user preferences, client-specific constraints), removes duplicates from the agent’s memory, and adds new practical rules drawn from accumulated experience.

2. Harvey’s performance: when the numbers speak

Harvey was one of the first companies to get access to Dreaming in a pilot version ahead of its public launch. This US company builds AI tools tailored for law firms, covering legal research, document drafting, and contract analysis.

Before Dreaming, Harvey’s agents had a well-documented problem: they forgot, from one session to the next, the quirks of certain file types, lawyers’ preferences, and mistakes to avoid. Every new file started nearly from scratch in terms of contextual learning.

After turning on Dreaming, Harvey measured roughly a 6x increase in its task completion rate. That gain came from two main factors:

• Agents asked far fewer clarifying questions, because they already knew the firm’s preferences and constraints.
• The number of recurring errors dropped considerably, since Dreaming had consolidated lessons from past sessions into the agent’s memory.

Harvey didn’t change its code and didn’t retrain its models. The agents improved on their own, between sessions, thanks to Dreaming’s automatic process. According to VentureBeat, which covered the launch, it’s precisely this autonomous, continuous quality that sets Dreaming apart from traditional improvement approaches.

3. How Dreaming works under the hood

For the more technically curious, here’s how Dreaming fits into the architecture of Claude agents. The agent has an external memory space, separate from the language model itself. That space stores information about past sessions: what worked, what failed, user preferences, and the quirks of each task.

Dreaming is a maintenance layer that runs periodically over this memory space. It performs three main operations:

• Consolidation: removing redundant entries and grouping similar information to avoid contradictions.
• Refresh: replacing outdated entries with more recent, relevant data, accounting for changing preferences.
• Pattern extraction: identifying recurring patterns that let the agent anticipate needs and avoid known mistakes before they happen again.

The result is an agent that gets progressively better at its job, with no human intervention and no retraining cost. It’s a different approach from traditional AI model improvement methods, which usually require data collection and training cycles that eat up time and resources.

Worth noting: Dreaming launched as a “research preview” at Code with Claude 2026. Access is by request, and Anthropic is positioning the feature as exploratory rather than production ready at this stage, with plans to widen access gradually.

4. The two other announcements alongside Dreaming

Dreaming didn’t arrive alone at Code with Claude 2026. Anthropic announced two other complementary features for its managed agents, which together form a coherent system:

• Outcomes: a structured feedback system after each task (success, failure, reason for failure), which feeds Dreaming quality data to work from. Without Outcomes, Dreaming wouldn’t be able to clearly tell what worked from what went wrong, which would limit how useful its adjustments could be.
• Multi-agent orchestration: agents can now delegate subtasks to other agents in parallel. A lead agent can orchestrate several specialized agents at once, which cuts down on bottlenecks for complex, multi-step projects.

Together, these three features address the main challenges of using AI agents at scale: persistent memory (Dreaming), the ability to evaluate one’s own performance (Outcomes), and efficiency on complex tasks (multi-agent orchestration). It’s an architecture designed to make agents more reliable and more autonomous the more they’re used.

5. Claude Code: the fastest-growing AI product in history

These new features arrive amid extraordinary growth for Anthropic, and for Claude Code in particular. Launched publicly in spring 2025, Claude Code has become, according to several industry sources, the fastest-growing software product in the industry’s history, outpacing records set by the biggest tech platforms in their early days.

The numbers tell the story: Claude Code now generates more than 2.5 billion dollars in annualized revenue and accounts for more than half of all enterprise spending across Anthropic’s platforms. In just one year, the product went from zero to a major player in AI-assisted development, with weekly active users doubling since the start of 2026.

More broadly, Anthropic’s annualized revenue jumped from 9 billion dollars in January 2026 to 47 billion in May 2026, according to information published by Fortune when the company confidentially filed its IPO prospectus on June 1, 2026. The company now counts more than 300,000 business customers, and the number of customers spending over 100,000 dollars a year has multiplied by 7 in a single year.

These numbers point to adoption that goes well beyond tech enthusiasts: businesses of every size, in every sector, are folding Claude into their daily workflows. AI agents are no longer a technological curiosity, they’re becoming a normal part of the workplace.

6. What this means for small businesses in Quebec

For a small business in Trois-Rivieres, in Mauricie, or in the Quebec City area, these advances aren’t yet a click away. Dreaming is in restricted access, and enterprise AI tools like Claude still need specialized guidance to be deployed properly and securely. But the trajectory is clear, and organizations that start preparing now will have a head start.

The sectors most likely to benefit from persistent AI agents in the near future are:

• Professional firms (law, accounting, consulting), where agents can retain client preferences and the specifics of each file from one meeting to the next.
• Automated customer service, where an agent that remembers past mistakes and customer habits can handle routine requests without escalating to a human.
• Document management, automated billing, and recurring report production, where repetitive formatting errors are especially costly in time.
• Cybersecurity and network monitoring, where persistent agents can better detect anomalies by drawing on a precise behavioral history.

The question is no longer “can AI do this?” The question is becoming “when do we start, and how?” Check out OKTO Solutions’ IT services to see how these technologies can fit into your organization in a structured way, adapted to your reality.

Frequently asked questions

What exactly is Claude Dreaming?

Claude Dreaming is an automatic process Anthropic launched in May 2026. It runs in the background while an AI agent is idle, reviewing past sessions to consolidate memory, remove redundancies, and extract useful patterns. The goal is for the agent to perform better with each new session, without anyone having to manually adjust its configuration.

Is Dreaming available to every business right now?

No. At launch on May 6, 2026, Dreaming was available only as a “research preview,” meaning by request for select partners. Anthropic plans to widen access gradually over the coming months. Interested businesses can submit an access request directly through the Claude.ai platform.

How are Claude agents different from a typical AI assistant?

A typical AI assistant answers one-off questions and generally has no memory between conversations. A Claude agent can carry out complex, multi-step tasks, use external tools, remember past interactions through persistent memory, and now improve between sessions thanks to Dreaming. It’s the difference between an advisor you consult once and a collaborator who learns and adapts over time.

A development worth watching closely for your business

Claude Dreaming marks a concrete step toward AI agents that no longer just answer questions, but actually improve over time, much like an employee who builds experience on the job. For Quebec small businesses, now is a good time to start exploring how these tools can fit into your operations, before your competitors get ahead. The team at OKTO Solutions can guide you from the initial audit through implementation, step by step and securely. Contact us to talk about your automation and AI needs.