Anthropic Is Teaching Claude to “Dream”: What the New Reflection Mode (and Orbit) Mean for Your Work

What does it mean for an AI to ‘dream’?

Let me get the framing out of the way. Claude is not actually dreaming. There is no consciousness here. “Dreaming” is the name Anthropic is using for a scheduled task where Claude reviews and consolidates its own stored memory between user sessions.[1]

Here is the plain-English version. When you use Claude with memory turned on, it stores facts about you: your role, projects, preferences, recurring themes. Over time that store gets messy. Duplicates. Outdated facts. Half-finished context. Just like your own notes.

The new dreaming feature runs a scheduled task on that memory: review what is in there, merge duplicates, surface patterns across past chats, prune what is no longer relevant.[1] It is a memory-hygiene loop. The first time AI is doing housekeeping on itself without you asking.

The closest human analogy is what your brain does in sleep: consolidate the day, drop what does not matter, integrate what does. Hence the name. It is metaphorical, not literal.

Why memory housekeeping is actually a big deal

If you have used ChatGPT or Claude memory for more than a few months, you have probably hit this problem. The AI “remembers” something stale. It still thinks you are working on a project you wrapped six months ago. It saves a duplicate of a fact you already gave it. It mixes up which client is which.

Memory drift is the dirty secret of AI tools. The feature looks impressive in the demo when it remembers your name. Twelve weeks in, your memory store looks like a junk drawer and the AI is making minor errors because of it.

The dreaming behaviour is Anthropic addressing this directly. The scheduled task reviews, deduplicates, and consolidates so the active context stays usable.[1]

For non-technical professionals, the practical implication is this. Memory is becoming the difference between an AI that feels like a tool and an AI that feels like a collaborator who knows your work. Tools with good memory hygiene compound in usefulness. Tools with bad memory get worse the longer you use them. Anthropic just made a meaningful bet on the first option.

Then there is Orbit: the proactive assistant

The second half of this story is Orbit. Orbit is Anthropic’s upcoming proactive assistant for Claude. It is currently in research preview, with references appearing in Claude’s mobile and desktop builds in early May 2026.[2]

Orbit pulls insights from a set of connected work apps and surfaces them proactively, without you prompting. The initial connector list is: Gmail, Slack, GitHub, Calendar, Google Drive, and Figma.[2][3] The pattern is similar to ChatGPT Pulse, but with a clearly different audience: where Pulse targets a broad professional set, Orbit is tilting toward people who work in collaborative tools (product managers, developers, designers, marketing teams).

The mechanic: Orbit watches the connected tools, runs scheduled briefings, and produces personalised cards. “What changed in this GitHub repo overnight,” “what was discussed in Slack while you were offline,” “which Figma frames got updated,” “what meetings are coming up.”[3]

Put dreaming and Orbit together and the picture sharpens. Dreaming makes Claude’s memory better. Orbit uses that improved memory to produce smarter proactive briefings. They are two parts of the same shift: from “AI you call” to “AI that calls you.”

Orbit vs ChatGPT Pulse: where they actually differ

If you read our guide to ChatGPT Pulse, you already know the proactive-AI playbook. Orbit covers the same territory with a different emphasis.

Connector mix. Pulse targets Gmail and Google Calendar at launch. Orbit goes wider out the gate: Gmail, Calendar, plus Slack, GitHub, Drive, and Figma.[2] The connector mix tells you who each product thinks its user is. Pulse is for executives and individual professionals. Orbit is for cross-functional teams who live in messaging and shared work tools.

Reflection layer. Pulse pulls fresh context each night and produces cards. Orbit is paired with dreaming: it reflects on what it has learned about you over time, not just yesterday. That should, in theory, produce smarter cards in week 8 than week 1.

Maturity. Pulse shipped to paying users in late 2025 and has had real production use.[4] Orbit is still in research preview as of May 2026.[2] The features sound similar; the polish gap is real.

Trust posture. Anthropic positions Claude differently from OpenAI on data handling. If your work involves regulated industries or sensitive client data, the privacy framing matters as much as the feature set. Read the policy on each, not the marketing.

The honest summary: these tools are going to get more similar than different. Pick the one that integrates best with the tools your team already uses, not the one with the slicker demo video.

Five practical use cases for non-technical professionals

The connector list (Slack, GitHub, Drive, Figma) makes Orbit sound like a developer tool. It is not only that. Here are five places it earns its keep for non-technical roles.

1. Cross-team status without nagging. If your work depends on three other teams hitting their deadlines, Orbit can surface what changed in their Drive folders or Slack channels overnight. You walk into your week knowing who is on track and who is not, without sending check-in messages.

2. Meeting prep that writes itself. You have a 9:30 with a client. Orbit produces a brief: the last Slack message they sent you, the file they edited yesterday, the open question from last week. The same brief you would have spent 20 minutes building, ready when you open your laptop.

3. Reviewing a teammate’s work without context-switching. If you manage someone who works in Figma or Drive, Orbit can summarise what they shipped overnight. You can give targeted feedback in 5 minutes instead of opening 14 tabs.

4. Catching up after time off. You took 4 days off. Normally that means a Monday morning of inbox triage. Orbit can produce a chronological brief of what happened across email, Slack, and shared docs while you were away. The first 30 minutes of your return become productive instead of overwhelming.

5. Pattern surveillance. Orbit’s reflection layer (the dreaming side) means it can flag when the same theme keeps coming up across tools. “You have mentioned X in 3 Slack channels this week. Want a summary?” That is the kind of insight you do not normally get without explicitly asking.

What to be honest about before you turn it on

Proactive AI sounds great until you think about the data flow.

Connector permissions are wide. Granting Orbit access to Gmail, Slack, Drive, and GitHub means an AI system is reading the content of those tools every day. That is a meaningful security decision. If you work somewhere with regulated client data, this needs IT and legal sign-off before it goes anywhere near your work account.

Memory accuracy still matters. Even with dreaming, memory drift happens. The model might “remember” something with confidence that is no longer true. The cleanup is helpful, not perfect. If a briefing card says “based on the open question with [Client X]” and that question was actually resolved last week, you should still verify before acting.

Information overload is a real risk. Proactive AI works against you if it produces too many cards. The first two weeks of using either Pulse or Orbit, expect noise. Use the thumbs-up / thumbs-down feedback aggressively, prune ruthlessly, and force yourself to spend less than 10 minutes on the morning briefing.

It is still early. Orbit is research preview. Dreaming is new. Both will have edge cases that surface only at scale. Treat 2026 as the year you learn how to work with proactive AI, not the year you bet your workflow on it.

For a deeper read on the limits of AI memory and reasoning generally, our guide to AI reasoning models covers what these systems actually do and do not understand.

What to do this week

Three concrete moves.

If you already use Claude: Open settings and check whether the dreaming toggle (memory cleanup) is on by default for your account. If not, turn it on. Spend 5 minutes reviewing what is in your memory store and delete anything that is stale.

If Orbit is available to you: Connect one tool to start (Calendar is the safest), not all six. Let it run for a week. Judge it on day 8, not day 1. Add a second connector only when the first one is producing useful cards.

If neither is available yet: Set up Claude memory properly with a seed prompt about your role, your projects, and the metrics you care about. The same setup that primes Pulse primes Orbit. By the time Orbit lands in your workspace, your account will be ready.

The bigger story is the shift itself. Across ChatGPT Pulse, Claude Orbit, and Gemini’s proactive features, the entire AI assistant category is moving in one direction: less prompting, more anticipation. The professionals who will get the most out of 2026 are the ones who learn to configure these systems well now, not the ones who wait until everyone is using them.