Your AI Is Dreaming About You: ChatGPT's Memory Overhaul and Who Controls the Record of Your Life

What OpenAI Shipped on June 4

When we wrote AI Memory Wars in April, the dividing line between AI memory systems was transparency. ChatGPT kept a partial list you could review. Claude kept human-readable files you could edit. Gemini kept an automatic profile you couldn't see at all. Two months later, OpenAI redrew the map.

On June 4, 2026, OpenAI announced "Dreaming: Better memory for a more helpful ChatGPT." The short version: the saved-memories list is no longer the primary way ChatGPT remembers you. In its place runs a background process, which OpenAI calls dreaming, that reads across your past conversations and continuously synthesizes an updated memory state. You don't ask it to remember things. It notices.

One detail in the announcement deserves more attention than it got. OpenAI said dreaming had already been running quietly for about a year, supplementing saved memories before anyone outside the company knew it had a name. The June launch wasn't the start of background synthesis. It was the moment OpenAI made it the architecture and stopped treating the list as the source of truth.

The rollout started with Plus and Pro subscribers in the US, with free and Go users plus more countries promised over the following weeks. That free-tier expansion is itself news: OpenAI says a roughly 5x reduction in the compute cost of running the dreaming process made it affordable to give persistent, synthesized memory to hundreds of millions of free accounts for the first time.

How Dreaming Actually Works

OpenAI hasn't published the architecture in detail, but the announcement and early coverage from Engadget and TechTimes sketch the mechanics clearly enough.

Between your conversations, a background process reviews what you've discussed, extracts durable context, merges it with what the system already believes about you, and rewrites the memory state. The result isn't a list of discrete facts. It's a synthesized model: your work, your projects, your gear, your preferences, your timeline. If you mentioned your camera in a conversation eight months ago, a product-recommendation chat today already knows what's compatible with your setup.

The most striking capability is temporal revision. Old saved memories went stale: ChatGPT would cheerfully reference a trip you took a year ago as upcoming. Dreaming updates entries as time passes, so "you're going to Singapore in July" becomes "you went to Singapore in July 2026" once the trip ends. The system doesn't just accumulate. It rewrites.

What you see is a memory summary page: a high-level overview of what ChatGPT has synthesized, with controls to correct details, dismiss items, and tell ChatGPT which topics to bring up or avoid. Note the word summary. You're reading a report generated about the memory, not the memory itself. We'll come back to why that distinction matters more than it sounds.

If you want the broader context on how ChatGPT, Claude, and Gemini got here, the prequel covers the full 2024 to early 2026 timeline. Everything below focuses on what changed since.

The Research Lineage: Models That Think While You Sleep

"Dreaming" sounds like marketing whimsy. It's actually a reasonably honest description of a research direction that's been building since 2023.

The lineage starts with MemGPT, the UC Berkeley project that treated an LLM like an operating system managing its own memory, paging information between a small working context and larger external storage. MemGPT became the startup Letta, founded by Charles Packer and Sarah Wooders out of Berkeley's Sky Computing Lab.

In April 2025, a Letta and UC Berkeley team (Kevin Lin, Charlie Snell, Yu Wang, Packer, Wooders, Ion Stoica, and Joseph E. Gonzalez) published "Sleep-time Compute: Beyond Inference Scaling at Test-time" (arXiv:2504.13171). The core idea: instead of doing all reasoning while the user waits, let the model think offline during idle time, pre-processing context and distilling raw history into useful form. In their experiments, this cut the test-time compute needed for the same accuracy by about 5x on one benchmark. The biological metaphor is deliberate: human memory consolidation happens substantially during sleep.

By 2026 the idea reached production at both major labs, under the same name. Anthropic shipped "dreaming" for Claude Managed Agents on May 6, 2026, as a scheduled process that reviews agent sessions, merges duplicates, prunes stale entries, and writes the distilled lessons into plain-text notes and playbooks. The legal AI company Harvey reported roughly 6x higher task-completion rates in tests after its agents could carry workarounds across sessions. OpenAI's consumer version followed four weeks later.

Here's the fork in the road, though. Anthropic's dreaming writes its conclusions to files humans can read and audit. OpenAI's dreaming writes to a memory state you see only through a summary. Same research lineage, same name, opposite answers to the question this article is about.

What Dreaming Genuinely Improves

Fairness first: the old system was bad, and this one is measurably better at its job.

OpenAI published an internal factual-recall evaluation showing task success at 41.5% with 2024's saved memories, 67.9% with the 2025 system, and 82.8% with Dreaming V3. On the same eval suite, preference adherence scored 71.3% and time-sensitive accuracy 75.1%. These are vendor numbers from an unreleased methodology, so treat them as directional rather than precise. But the direction matches what users complained about for two years: saved memories were stale, redundant, and weirdly selective.

The concrete improvements are easy to list. Stale facts get revised instead of contradicting the present. Context you mentioned naturally in conversation gets captured without the "remember this" ritual. Relevant details surface across topic boundaries, so your photography setup informs your shopping question. And the 5x compute reduction brings all of this to free accounts, which previously got the weakest memory features.

There's also an honest argument that synthesis is simply what memory means. Human memory isn't an append-only log either. We consolidate, revise, and forget, and that's most of why remembering works. So the engineering is defensible and the benefits are real. The problem isn't what dreaming does. It's what you can no longer do.

What It Costs: Audit, Contestability, Lock-In

The old saved-memories list was clunky, but it had a property nobody priced correctly until it was gone: it was enumerable. You could read every item, point at one, and delete it. The model's beliefs about you and the list you audited were the same object.

Dreaming breaks that identity. Three specific losses follow.

Audit. The memory summary page is a generated overview, and OpenAI's own documentation acknowledges it doesn't necessarily include everything ChatGPT may remember from past conversations. You are auditing a report about the record, not the record. Coverage from TechTimes flagged this as the central change: the personalization engine got rewritten, and the audit trail got narrower at the same time.

Contestability. Suppose the synthesis draws a wrong inference about your health, your politics, or your employer. With a list, you delete the line. With a synthesized state, the controls are softer: per TechTimes, selecting "don't mention this again" reduces future references to a detail but doesn't delete the underlying entry, and deleting a conversation doesn't remove memories already derived from it. You have to hunt down both the inference and its source, and you can't verify the hunt succeeded. A system that quietly revises itself can also quietly revise a correction you never got to review.

Lock-in. This is the slow cost. The prequel celebrated March 2026's portability moves, when Claude and Gemini shipped import tools that could ingest competitor exports. Those tools work on legible exports. A continuously synthesized memory state is exactly the kind of artifact that exports badly: what would you even export, the summary? Every month of dreaming makes your ChatGPT measurably more useful to you and your context less reconstructible anywhere else. We've written before about why personal context is becoming the real moat, and Dreaming is the cleanest example yet: the moat is now built automatically, in the background, on the vendor's side of the fence.

None of this requires bad intent from OpenAI. It's a structural property of the design. When the record of you is an artifact only the vendor's process can write, your relationship to that record changes from owner to subject.

ChatGPT vs Claude vs Gemini: Memory in June 2026

The prequel's comparison table is now two product generations old, so here's the updated landscape. Since March 2026: Google launched Personal Intelligence (January 14, 2026 for AI Pro and Ultra in the US, expanding to free users from late March), which pulls context from Gmail, Photos, Search, and YouTube history. Anthropic brought memory to all Claude plans on March 2, 2026, and added agent dreaming in May. And OpenAI shipped the overhaul this article is about.

The pattern worth noticing: the three vendors are no longer converging. Claude doubled down on legibility, Google doubled down on breadth (your inbox and photo library are now memory), and OpenAI doubled down on quality of synthesis at the expense of inspection. Pick your tradeoff, because you're definitely making one.

Developer Sidebar: The Agent-Memory Stack

If you build with LLMs rather than just chat with them, there's a parallel memory ecosystem worth knowing, and it has mostly taken the opposite position on ownership.

Letta (the MemGPT team) treats memory as an agent-managed hierarchy and runs sleep-time compute as a first-class feature, with agents that reorganize memory between sessions. Zep builds memory as a temporal knowledge graph through its open-source Graphiti library (27k GitHub stars), where every fact carries valid-at and invalid-at timestamps, so "user moved to Tokyo" supersedes rather than contradicts "user lives in Berlin." That's the same temporal-revision trick as Dreaming, except queryable and inspectable. Mem0 raised a $24M Series A in October 2025 to be the managed memory layer for AI apps, crossed 41,000 GitHub stars at the announcement (it's above 58,000 now), and became the memory provider for AWS's Agent SDK. LangMem is LangChain's take, splitting memory into semantic, episodic, and procedural types.

Anthropic's API-side answer is the memory tool (memory_20250818), which gives Claude a persistent file directory it can read and write across sessions. The notable design choice: the files live client-side, on your infrastructure. Anthropic's model does the remembering; you hold the memories.

Evaluation has standardized around two benchmarks. LoCoMo tests very long-term conversational memory across multi-session dialogues, including temporal-reasoning and multi-hop questions. LongMemEval pairs each question with roughly 115k tokens of chat history in a needle-in-a-haystack setup. Vendor leaderboard claims are marketing-heavy (everyone is state of the art on the benchmark they chose), but the benchmarks themselves are how you'd verify any memory claim, including OpenAI's, if the eval were public. It isn't.

The irony is hard to miss. The developer stack, built by people who have to debug these systems, converged on memory as legible, timestamped, exportable data under the application's control. The largest consumer AI product just moved its 800 million users the other way.

The Memory Layer You Can Actually Own

So what do you do, practically, if you want the benefits of AI memory without making a vendor's background process the authoritative record of your life?

You keep a record the vendor doesn't write. The prequel called this user-owned memory, and Dreaming makes the case sharper rather than weaker. A synthesized memory state is the vendor's interpretation of you. Your highlights, notes, and annotations are your interpretation of you, and the difference between those two artifacts is the difference between a credit score and a diary.

This is the design premise behind Glasp's web highlighter: as you read, you save the passages that struck you and the notes you attached to them, in plain text, under your account, in a form you can export at any time. Nothing infers. Nothing rewrites overnight. The record grows only when you act, which means it stays contestable in the simplest possible way: you wrote it, so you can change it.

The two layers aren't rivals; they compose. When you ask Glasp's AI chat a question, it reads from highlights you explicitly chose to keep, which is the consent model the big platforms abandoned. And because the record is portable text rather than an opaque synthesized state, it works as a personal RAG corpus for any model you point at it: ChatGPT today, Claude tomorrow, whatever ships in 2027. We've laid out the operational side of this in our guide to managing personal context in an AI-first workflow.

Use Dreaming. It's genuinely good at what it does. Just don't let it be the only record. Convenience memory should live with the vendor; the canonical copy of what you've read, thought, and decided should live somewhere you hold the keys.

Frequently Asked Questions

How does ChatGPT memory work after the Dreaming update?

A background process now runs between your conversations, reading past chats and synthesizing a continuously updated memory state. Memories revise themselves over time (a planned trip becomes a past trip), and context you mention naturally gets captured without you asking. You review it through a memory summary page in settings.

Did ChatGPT remove the saved-memories list?

The list is no longer the primary mechanism. OpenAI's announcement describes the new system as built on dreaming, with a memory summary page replacing list management for most purposes. You can still add details, correct them, and dismiss items from the summary, but the underlying memory is a synthesized state rather than an enumerable list of facts.

Can I see everything ChatGPT remembers about me?

No. The memory summary is a high-level overview, and OpenAI's documentation acknowledges it doesn't necessarily include everything ChatGPT may remember from past conversations. Also note that deleting a conversation doesn't remove memories already derived from it, and "don't mention this again" suppresses a detail without deleting it. For sensitive topics, use Temporary Chat, which keeps the session out of memory.

Is Dreaming available on the free tier?

It's rolling out. The June 4, 2026 launch covered Plus and Pro users in the US first, with free and Go accounts and more countries promised over the following weeks. OpenAI says a roughly 5x cut in the compute cost of the dreaming process is what made free-tier synthesis affordable for the first time.

Are the 82.8% recall numbers trustworthy?

They're from OpenAI's internal evaluation, with no published methodology, so treat them as directional. The reported progression (41.5% in 2024, 67.9% in 2025, 82.8% with Dreaming V3) matches the widely shared experience that the old system was stale and unreliable. Public benchmarks like LoCoMo and LongMemEval exist for exactly this kind of claim, but OpenAI hasn't reported against them.

Conclusion

ChatGPT's memory used to be a list. Now it's a process. That single change explains both why the product got better and why the question in this article's title got urgent. A list can be read, contested, and exported. A process can only be trusted.

The June 2026 landscape gives you three vendor philosophies: OpenAI's high-quality opaque synthesis, Claude's legible files, Gemini's ecosystem-wide reach. All three are useful. None of them is yours. The memory that's yours is the one you author: the passages you chose to keep, the notes you wrote in the margin, the record that doesn't change while you sleep.

Start building that record now, while your AI dossiers are still young. Install Glasp's web highlighter, highlight what actually matters to you this week, and you'll have begun the one memory layer that no background process can rewrite and no terms-of-service update can take. Your AI is already dreaming about you. Make sure you're keeping notes of your own.