ChatGPT: Enhanced Memory for Richer Interactions

⚡

Key Takeaways

1OpenAI introduces a new memory architecture for ChatGPT, aimed at addressing stale and scalability issues.

2The update is initially available for Plus and Pro users in the United States, with a global rollout planned.

3The enhanced memory allows ChatGPT to retain user preferences and projects, making interactions more personalized.

💡Why it matters — This evolution strengthens ChatGPT's ability to provide smoother and more relevant interactions, thereby increasing its usefulness for regular users.

A More Efficient Memory System for ChatGPT

OpenAI has begun rolling out an enhanced memory system for ChatGPT, designed to overcome challenges related to obsolescence, correction, and scalability. This new system aims to provide a smoother experience for the hundreds of millions of users interacting with ChatGPT over the years.

Memory plays a crucial role by allowing ChatGPT to remember users' preferences, projects, and constraints. This means that future conversations can start with shared context, rather than starting from scratch each time. For the past two years, memory has become an essential component of the user experience, helping ChatGPT better understand context and achieve more meaningful goals over time.

This update is currently available for Plus and Pro users in the United States. OpenAI plans to expand it to other countries and to users of the Free and Go versions in the coming weeks.

History of Memory in ChatGPT

The memory feature was first introduced in April 2024, under the name "saved memories." This feature allowed users to ask ChatGPT to remember specific information for future conversations. However, saved memories required explicit signals to be triggered, such as a clear instruction to remember a trip to Singapore in July. In practice, this resembled a person taking notes but forgetting everything that wasn't written down. Additionally, these memories tended to become outdated over time, losing their relevance.

In April 2025, OpenAI updated ChatGPT's memory by introducing the first version of the "dream." This method allows ChatGPT to automatically synthesize memories in the background, referring to the history of discussions. Unlike saved memories, the dream operates in the background, enabling ChatGPT to learn from numerous conversations and provide always fresh and relevant context. The dream also includes contexts that naturally occur in conversation, without relying on explicit requests.

Over the past year, the dream has complemented saved memories, significantly improving ChatGPT's ability to personalize responses and compensate for the obsolescence of saved memories. However, it was not sufficient as a standalone memory system.

Today, OpenAI introduces a significantly more efficient and computationally economical memory architecture built on the dream. Memories synthesized by the dream are searchable via a visible summary on the memory summary page. From this summary, users can quickly grasp the key points of what ChatGPT knows about them, add or update information, and provide instructions on topics that ChatGPT should address and when. To delve deeper into a particular area, users simply need to discuss it with the model.

Evaluating Memory

To determine what constitutes "good memory" in ChatGPT, several criteria are considered:

Conveying useful context: When a user shares information with ChatGPT, it should remember it for future discussions.
Tracking preferences and constraints: If a user expresses a preference, such as being vegetarian, ChatGPT should act in accordance with that preference in the future.
Staying current over time: Memory must take the passage of time into account. For example, if a user is planning a birthday party for next Saturday, ChatGPT should know that this event has passed once Sunday arrives.

OpenAI evaluates how the memory of ChatGPT Plus and Pro has improved over time against these goals. This assessment has been conducted each year since the introduction of saved memories in 2024, through the addition of the dream in 2025, up to the current version of the dream in 2026.

Context Transmission in Conversations

With the enhanced memory, when you start a new conversation with ChatGPT, you don't need to reintroduce yourself. ChatGPT can rely on previous context, which is particularly useful for complex and long-term projects.

For example, if you are using ChatGPT to buy new photo equipment compatible with your current device, and you have already discussed your photo setup with it, you can ask for products compatible with "my photo setup" and receive recommendations tailored to your needs.

Example Without Memory

When a user asks, "What should I buy to use TTL for my underwater photography setup?", ChatGPT must provide a basic shopping checklist, explaining the differences between fiber optic TTL and electrical TTL.

Example With Memory

With memory, ChatGPT can check the user's actual setup, for instance, "Sony A1 II in Nauticam NA-A1II, with Backscatter Mini Flash 3 and Inon Z-330 strobes," and provide specific recommendations based on that setup.

Thus, for the mentioned setup, ChatGPT might recommend the Backscatter Smart Control TTL LED Nauticam Flash Trigger for Sony for the Mini Flash 3, or the UW Technics 11075-HSS TTL Converter for Sony/Nauticam for the Inon Z-330.

In conclusion, this new memory architecture allows ChatGPT to provide more relevant and personalized responses, thereby enhancing the overall user experience.