Claude Can Now Dream Between Sessions, and Here Is What That Means for L&D

So Claude Can Sleep Now

Anthropic announced something at their Code with Claude conference on May 6 that I have been thinking about ever since. It is called Dreaming, and it does roughly what the name suggests. Between active sessions, Claude reviews what happened, pulls out the patterns that mattered, and writes a cleaner version of what it learned into its memory.

Then the next time you (or your learners, or your team) start a session, the agent picks up where the smarter version of itself left off.

For most software features, I would not lead with the biology metaphor. For this one, the biology metaphor is the whole point. Anthropic actually built it in two phases that map to what happens in a human brain during sleep, and once you see what it is doing, it is hard to unsee the implications for anyone designing learning experiences.

What Dreaming Actually Is

Here is the simple version. Every time you use an AI agent, the session generates a transcript: what you asked, what it did, what worked, what failed. Most of that gets thrown away when the session ends. Dreaming changes that. It runs in the background between sessions and does two things.

Importantly, Dreaming does not touch the original session data. It builds a new, cleaner memory layer alongside it. The raw transcripts stay intact, and the consolidated memory is what gets loaded at the start of the next session.

Why That Brain Analogy Matters

I taught for years before I moved into instructional design, and one of the things you notice early as a teacher is that your students do not become better learners by sitting in your class. They become better learners by sleeping on what they learned in your class. The work the brain does between lessons is where the real consolidation happens.

Dreaming is the first AI feature I have seen that takes that idea seriously at the architectural level. The agent is not just storing more raw data. It is doing what a learner does: looking back, noticing patterns, throwing out noise, and updating its mental model.

For anyone who has ever struggled with an AI tool that keeps making the same mistake or forgetting context, this is the feature that, in theory, fixes that.

How It Actually Runs

Dreaming runs asynchronously on a schedule. It does not slow down your active sessions. The flow looks like this:

1. Session logging. Every interaction generates a transcript with inputs, outputs, tool calls, reasoning, and outcomes.
2. Pattern extraction. Between sessions, Dreaming runs meta-reasoning across batches of those logs to identify what worked, what failed, and what is recurring.
3. Memory consolidation. The extracted insights get written into structured memory as updated preference profiles, decision rules, curated knowledge snippets, and flagged failure patterns.
4. Session-start injection. The next session loads from the consolidated memory, not the raw logs.

The result is an agent whose memory gets cleaner and more useful over time instead of bloating and degrading.

Two Modes: Automatic vs Human Review

This is the part that matters most for L&D, especially anyone working in a regulated or compliance-heavy space.

For most L&D production deployments, Human Review is the right starting point. You get the pattern recognition without giving the agent unilateral write access to its own brain. If you work in healthcare, financial services, or any regulated training space, this is not optional. The agent should not silently update what it thinks the right answer is.

The /dream Command in Claude Code

If you are already using Claude Code (which is the tool I use to build basically everything on this site), there is a practical implementation called AutoDream that you can use right now. It runs as a background sub-agent that reads your existing CLAUDE.md memory files, prunes stale entries, merges duplicates, resolves contradictions, and produces a cleaner version of your memory file.

You can also trigger it manually with the /dream command.

If you have been building something in Claude Code for weeks or months, you have probably noticed that your CLAUDE.md file slowly turns into a chaotic pile of contradictory instructions, outdated context, and notes you forgot you wrote. /dream cleans that up automatically. For anyone running long projects, this alone is worth knowing about.

Why L&D Should Pay Attention

This is the section I have been mentally rewriting since the announcement. Here is the thing. The core problem Dreaming solves is exactly the problem that breaks long-running L&D workflows.

Training programs evolve over months. Learner cohorts change. SME input contradicts earlier drafts. LMS data reveals what actually worked and what did not. Right now, an instructional designer is the one carrying all of that institutional memory around in their head, re-briefing every tool, every contractor, every new AI session. Dreaming is the first feature I have seen that lets an agent do some of that carrying for you.

Adaptive AI Tutors That Actually Improve

This is the biggest one. Imagine an AI tutor built into a course. Without Dreaming, every learner session is a fresh start, and the tutor never gets sharper at teaching your specific content. With Dreaming, across hundreds or thousands of learner sessions, the tutor learns:

• Which explanations consistently caused confusion and which landed
• Which quiz questions learners reliably miss (which usually means your content has a gap, not that your learners are bad)
• Which learner archetypes need different pacing, scaffolding, or vocabulary
• Which analogies and examples actually stick

Over weeks, the tutor becomes measurably better at teaching your content to your audience. Without you changing a single prompt.

Curriculum QA That Remembers Your Standards

If you have ever used Claude or any AI to review course drafts, check alignment to learning objectives, or flag SME inconsistencies, you know the painful part. Re-briefing the agent every single time. "We use Bloom's at the application level." "This client hates passive voice." "These two modules contradict each other on the definition of competency."

A Dreaming-enabled QA agent consolidates those standards and carries them forward. You stop being the only keeper of every brand voice rule and content standard.

SME Knowledge Capture Without Re-Asking the Same Questions

If you have run SME intake sessions, you know they are messy. Experts contradict themselves. They add context weeks later. They forget what they said in the first meeting. A Dreaming-enabled intake agent can consolidate input across multiple sessions, flag contradictions between what different experts said, build a growing clean knowledge base of the domain, and identify what is still missing.

For anyone doing technical or compliance training that lives or dies on getting SMEs right, this is huge.

Kirkpatrick-Aligned Evaluation Agents

For L&D teams tracking training effectiveness, an agent ingesting Level 1 through 4 data (reaction surveys, assessments, behavior change reports, business results) can use Dreaming to build a cross-program picture over time. It learns which interventions correlated with behavior change, which evaluation instruments were most predictive, and which business units had the highest transfer rates. That institutional knowledge starts compounding instead of getting lost when someone leaves the team.

Onboarding That Self-Improves With Every Cohort

New hire onboarding agents can use Dreaming to track where new hires drop off, which FAQs get asked repeatedly (revealing content gaps), and which scenario branches lead to faster competency. The agent improves the onboarding experience as it runs more cohorts. That is closer to a continuously optimized onboarding curriculum than anything I have seen built without a full team behind it.

The Outcomes Feature Pairs With This (And It Matters)

Dreaming pairs with another Anthropic feature called Outcomes, which is in public beta. The short version: you define a rubric, a separate Claude instance grades the agent's output against that rubric, and Dreaming then consolidates the graded successes and failures into future memory.

For L&D, the rubric possibilities map almost perfectly to what we already measure:

• Did the learner correctly apply the concept in a scenario?
• Did the generated content hit the right Bloom's level?
• Did the agent's feedback to a learner actually improve comprehension on the next attempt?

Use Cases for Beginners and Everyday Users

You do not need to be a developer or run a research preview to start working with the ideas behind Dreaming. Here is what is actually accessible to regular users right now.

Personal Learning Coach

Start a Claude Project for whatever you are learning (Spanish, public speaking, prompt engineering). After each session, ask Claude to update a running learning profile note: what you have mastered, where you keep struggling, what explanations clicked. Over weeks, you build a picture of yourself as a learner, and Claude stops re-explaining things you already know.

Writing Partner That Learns Your Voice

Use a Claude Project for all your writing (blog posts, emails, proposals). Periodically tell Claude what feedback you got, what tones landed, what phrases feel off-brand. A Dreaming-enabled agent will eventually do this automatically. For now, doing it manually trains you to be more intentional about voice, which is honestly the bigger win anyway.

Study and Certification Prep

Students prepping for certifications (ATD CPTD, AWS certs, PMP, whatever you are working toward) can run a Claude project across dozens of practice sessions. Track which topic areas you consistently miss, which question formats trip you up, which explanations finally made things click. Update the project notes after each session. You are essentially doing manual Dreaming.

Small Business FAQ Bot

A small business owner can deploy a Claude-powered bot through no-code tools like MindStudio for FAQs, booking questions, or customer support. With Dreaming, the bot learns which questions repeat (revealing content gaps), which responses frustrate people, and which approaches actually convert.

Claude Code Users: Just Run /dream

If you are using Claude Code on any kind of ongoing project (a website, a curriculum, a research workflow), run /dream on it. It cleans up your CLAUDE.md file in seconds. This is the easiest immediate win available.

What to Actually Watch Out For

I want to be honest about the limits, because the announcement coverage has been a little breathless.

• Memory quality mirrors session quality. Vague, scattered, or poorly structured sessions produce weak consolidations. Junk in, organized junk out.
• These are inferences, not facts. The patterns Dreaming extracts are probabilistic. The agent can over-generalize from a small number of sessions. If a tutor saw three learners stumble on Module 4 and concluded "everyone struggles with this content," that might be wrong.
• Bad habits consolidate too. If your agent spent weeks doing something poorly, Dreaming will organize that bad approach more efficiently. Garbage in, neatly-filed garbage out.
• Memory lock-in. Consolidated memories currently live on Anthropic's infrastructure. The learning does not travel with you to other platforms. If you are building something long-term, factor that in.
• Learner data privacy. Dreaming consolidates behavioral data from sessions. Your data governance policy needs to account for what learner interaction data ends up in the memory layer. This matters especially for K-12, healthcare training, or anything under FERPA, HIPAA, or GDPR.
• Multi-agent oversight gets complex. If you have a system with multiple dreaming agents, tracking why an agent's behavior changed between sessions can get fuzzy fast.

Pricing and Access

Dreaming lives inside Claude Managed Agents, which runs at standard Claude API token rates plus $0.08 per active session hour. Idle time is free, and Dreaming, Outcomes, and Webhooks do not carry separate per-feature charges on top of that.

It is currently in research preview, which means you have to apply for developer access rather than flipping a switch. If you are building something serious for your organization, that application is worth submitting now.

How to Get Started Right Now

Even before full Dreaming access opens up beyond research preview, here is what I would actually do this week:

1. Pick one ongoing AI project (your blog drafts, your course outlines, your daily standup notes) and start treating its memory as something you intentionally curate.
2. Open a Claude Project for it and write a structured memory document. What does success look like? What are the constraints? What feedback have you already given that you keep having to repeat?
3. Run a manual consolidation session every few weeks. Ask Claude to review your project notes and propose updates. You are essentially DIY-Dreaming.
4. If you are in Claude Code, try /dream. It is free, it is fast, and it will probably surprise you with how much cruft has built up in your CLAUDE.md.
5. If you are building something serious, apply for Managed Agents research preview. The link is on Anthropic's developer site.

The Bottom Line

For most consumer AI tools, "memory" has meant "we save your past chats so you can scroll back through them." Dreaming is the first feature I have seen that treats memory the way teachers and instructional designers have always thought about it. As the thing the learner does between lessons, not the thing that lives on a server.

For L&D specifically, this matters because we have been waiting for AI tutors and content tools that actually get better over time at teaching our specific content to our specific learners. Dreaming is the closest thing to that I have seen on the roadmap. It is not magic, it has real limits, and you still have to be intentional about what you feed it. But it is a real shift, and worth understanding before it shows up in the tools you are already using.