OpenAI is rolling out a more capable and scalable memory synthesis system called "dreaming," designed to address staleness, correctness, and scalability issues that arise when memory serves hundreds of millions of users over multi-year timeframes in ChatGPT.
Memory helps ChatGPT learn user preferences, projects, and constraints so that future conversations can build on shared context rather than start from scratch.
Over the past two years, memory has become a critical part of the ChatGPT experience, enabling the assistant to better understand user context and help accomplish meaningful goals over time. This is central to OpenAI's goal of making ChatGPT more useful: knowing users, helping them, and doing more for them.
This update is available to Plus and Pro users in the US initially, with a rollout to additional countries and Free and Go users planned over the coming weeks.
How Memory Has Evolved
Memory originally launched in April 2024 as "saved memories," allowing users to ask ChatGPT to remember information and carry it into future chats.
Saved memories were only written during conversations and relied on strong cues to trigger, such as an explicit instruction like "remember I'm traveling to Singapore in July." In practice, interacting with this system felt like talking to someone who took a few notes but forgot everything else. Saved memories also tended to go stale over time, eventually becoming incorrect or irrelevant.
In April 2025, OpenAI updated ChatGPT's memory by giving the model the ability to reference chat context beyond the saved memories list. This was accomplished by introducing the first version of dreaming - a method for ChatGPT to automatically curate memories in the background by referencing chat history.
Unlike saved memories, dreaming uses a background process that allows ChatGPT to learn from many conversations and synthesize its memory state, always providing the freshest, most relevant context. Dreaming also makes it easier for memory to capture context that arises naturally in conversation, without requiring explicit "remember this" requests.
Over the past year, dreaming supplemented saved memories to create a significant improvement in ChatGPT's ability to personalize responses and counteract memory staleness. However, it historically was never sufficient as a standalone memory system.
With today's launch, OpenAI is introducing a significantly more capable and compute-efficient memory architecture built on top of dreaming.
The memories synthesized by dreaming are reviewable through a summary visible on a memory summary page. From this summary, users can quickly see the highlights of what ChatGPT knows about them, add or update personal information, and provide instructions on what topics ChatGPT should raise and when. Users who want to drill into a specific area can simply chat with the model.
How OpenAI Evaluates Memory
When considering what "good memory" looks like in ChatGPT, several criteria stand out:
- Carry forward useful context: A user tells ChatGPT something once, and it remembers that information in subsequent chats.
- Follow preferences and constraints: If a user describes a preference (e.g., being vegetarian), ChatGPT should take actions consistent with that preference going forward.
- Stay current over time: Memory should account for the passage of time. For example, "The user is planning their birthday party for next Saturday" should eventually be updated once that Saturday has passed.
OpenAI evaluated how ChatGPT Plus and Pro memory has improved over time across these three objectives, comparing:
- 2024: Saved memories
- 2025: Saved memories + Dreaming V0
- 2026: Dreaming V3
Carrying Forward Context
When users start a new chat with ChatGPT, they no longer need to introduce themselves from scratch. ChatGPT can save time and build on prior context, which is especially valuable for complex, long-running projects.
For example, if a user has previously discussed their camera setup with ChatGPT, they can ask for products compatible with "my photography setup" and receive tailored recommendations that match their specific gear - rather than a generic checklist requiring the user to do complicated compatibility research on their own.
In OpenAI's evaluations, the new dreaming-based system improves the model's ability to recall relevant facts from prior conversations.
Following Preferences
Memory also helps ChatGPT respond in ways that better match user preferences and constraints.
For instance, if a user is planning a trip to Singapore and ChatGPT already knows from past travel planning that they enjoy wildlife photography, prefer hotels with strong air conditioning, and favor quiet dinners over crowded bars, the resulting itinerary will be personalized accordingly - rather than producing a generic tourist guide.
Preferences can take several forms:
- Instructions for how ChatGPT should respond (e.g., "don't bring up Stan again")
- Personal preferences or constraints (e.g., "I'm vegetarian")
- Implicit preferences that shape relevance (e.g., "I live near San Francisco" leads to locally tailored options)
In developing the new memory system, OpenAI improved ChatGPT's ability to apply relevant preferences from past conversations.
Staying Current Over Time
Traditional memory systems can become stale. For example, a user tells ChatGPT "I'm in Singapore and need a dinner recommendation for tonight." Time passes, the trip ends, but ChatGPT still thinks the user is in Singapore.
With dreaming, memories are automatically updated as time passes. ChatGPT can revise its memory from "You're going to Singapore in July" to "You went to Singapore in July 2026" once the trip concludes. When the user is back home, ChatGPT can again provide recommendations tailored to their home location and time zone.
In OpenAI's memory evaluations measuring whether ChatGPT correctly responds to prompts where the passage of time materially affects the answer, dreaming provides a substantial improvement.
A More Scalable Foundation for the Future
While dreaming-based memory has been available to Plus and Pro users for some time, OpenAI is now able to offer Free users a version that meets their quality bar and is practical to serve at scale. Recent improvements reduced the compute required to serve dreaming to Free users by approximately 5x, making it possible to begin rolling it out to Free users over the coming weeks and to increase memory capacity for Plus and Pro users.
Looking ahead, dreaming now provides OpenAI with a shared memory foundation for all users. This update represents OpenAI's most capable memory system to date, and the team plans to continue improving it.