From saved facts to a living narrative

The original ChatGPT memory, launched in April 2024, relied primarily on explicit “saved memories”—users had to tell the model what to remember . Dreaming V1, introduced in April 2025, added a background process that could reference chat history automatically, but it remained supplemental to the saved-memory list .

Dreaming V3 is a fundamental rearchitecture. Instead of storing disconnected bullet points (e.g., “user is a software engineer,” “user has a dog named Max”), the system now categorizes information—occupation, hobbies, travel patterns, preferences, constraints—and compresses it into cohesive narrative profiles . This is not just a data-structure change; it’s a semantic shift. The model understands you as a person with a story, not a collection of trivia.

The synthesis runs as a background job that periodically reviews chat history, extracts relevant signals, and updates the profile automatically . Users no longer need to say “remember this.” The system observes, synthesizes, and refreshes its understanding as your life and conversations evolve .

What makes a good memory?

OpenAI defines three quality criteria for a useful memory :

1. It carries useful context across conversations.
2. It respects user preferences and stated constraints.
3. It updates itself over time without user intervention.

These aren’t aspirational goals—they’re the evaluation framework that produced the headline accuracy number. Under the new system, factual accuracy jumped from 52.2% to 75.1%, a statistically significant improvement that reflects fewer stale, contradictory, or outright incorrect memories persisting in the background .

User-facing features

Alongside the architectural overhaul, OpenAI shipped several user-facing improvements:

• Memory summary page: A new interface shows a digestible overview of what ChatGPT knows about you. Users can view highlights, manually add or update personal details, and set instructions about what topics ChatGPT should raise (or avoid) and when .
• Automatic pruning and refresh: Outdated memories are actively removed, and new context from recent chats is synthesized without user prompts. This keeps responses aligned with your current reality—not the version of you from six months ago .
• Cross-context awareness: ChatGPT can now pull relevant context from past chats, saved memories, uploaded files, and connected apps to deliver responses that reflect the full scope of your interactions with the platform .

The economics: why Dreaming V3 unlocks the free tier

Previous memory architectures were computationally expensive to run across hundreds of millions of users. Synthesizing memory at scale required substantial inference resources, which made it economically viable only for paying subscribers.

Dreaming V3 changes that equation. OpenAI reports that the computational cost of memory synthesis has been reduced to roughly one-fifth of its previous level . This 5x efficiency gain is the strategic enabler: it makes personalized, stateful AI affordable for free-tier users without eroding margins. As a result, the same sophisticated memory system shipping to $200/month Pro users will also reach users on the free plan over the coming weeks .

The competitive landscape: everyone wants an AI that remembers

OpenAI’s move is part of an industry-wide inflection point. By mid-2026, all three major consumer AI labs have shipped persistent memory systems :

• Anthropic launched Claude Dreaming on May 6, 2026, using offline curation that reviews up to 100 past sessions to identify patterns and self-improve.
• Google introduced Gemini Spark’s account-level context at I/O on May 19, 2026, embedding memory directly into its assistant ecosystem.
• OpenAI now matches and, in some respects, leapfrogs both with a system that synthesizes narrative profiles in the background for hundreds of millions of users.

This is not a niche feature race. Persistent memory represents a fundamental shift from stateless to stateful AI—from tools that answer questions to assistants that build an ongoing relationship with each user .

Strategic significance: the personalized assistant vision

Dreaming V3 is one of the clearest signals yet that OpenAI is building toward a proactive, context-aware personal assistant rather than a general-purpose chatbot . By reducing reliance on explicit commands and instead inferring context automatically, the system approaches a long-standing vision: an AI that knows your work projects, your goals, your dietary restrictions, your travel plans, and your communication style—without being retold each session.

The combination of dramatically lower compute costs and narrative-level understanding makes this vision economically viable at global scale. For OpenAI, personalized memory isn’t just a feature—it’s the foundation for agent-like behavior, where an AI can act on your behalf with genuine understanding of who you are and what you need .

Access and rollout

From saved facts to a living narrative

The original ChatGPT memory, launched in April 2024, relied primarily on explicit “saved memories”—users had to tell the model what to remember . Dreaming V1, introduced in April 2025, added a background process that could reference chat history automatically, but it remained supplemental to the saved-memory list .

Dreaming V3 is a fundamental rearchitecture. Instead of storing disconnected bullet points (e.g., “user is a software engineer,” “user has a dog named Max”), the system now categorizes information—occupation, hobbies, travel patterns, preferences, constraints—and compresses it into cohesive narrative profiles . This is not just a data-structure change; it’s a semantic shift. The model understands you as a person with a story, not a collection of trivia.

The synthesis runs as a background job that periodically reviews chat history, extracts relevant signals, and updates the profile automatically . Users no longer need to say “remember this.” The system observes, synthesizes, and refreshes its understanding as your life and conversations evolve .

What makes a good memory?

OpenAI defines three quality criteria for a useful memory :

1. It carries useful context across conversations.
2. It respects user preferences and stated constraints.
3. It updates itself over time without user intervention.

These aren’t aspirational goals—they’re the evaluation framework that produced the headline accuracy number. Under the new system, factual accuracy jumped from 52.2% to 75.1%, a statistically significant improvement that reflects fewer stale, contradictory, or outright incorrect memories persisting in the background .

User-facing features

Alongside the architectural overhaul, OpenAI shipped several user-facing improvements:

• Memory summary page: A new interface shows a digestible overview of what ChatGPT knows about you. Users can view highlights, manually add or update personal details, and set instructions about what topics ChatGPT should raise (or avoid) and when .
• Automatic pruning and refresh: Outdated memories are actively removed, and new context from recent chats is synthesized without user prompts. This keeps responses aligned with your current reality—not the version of you from six months ago .
• Cross-context awareness: ChatGPT can now pull relevant context from past chats, saved memories, uploaded files, and connected apps to deliver responses that reflect the full scope of your interactions with the platform .

The economics: why Dreaming V3 unlocks the free tier

Previous memory architectures were computationally expensive to run across hundreds of millions of users. Synthesizing memory at scale required substantial inference resources, which made it economically viable only for paying subscribers.

Dreaming V3 changes that equation. OpenAI reports that the computational cost of memory synthesis has been reduced to roughly one-fifth of its previous level . This 5x efficiency gain is the strategic enabler: it makes personalized, stateful AI affordable for free-tier users without eroding margins. As a result, the same sophisticated memory system shipping to $200/month Pro users will also reach users on the free plan over the coming weeks .

The competitive landscape: everyone wants an AI that remembers

OpenAI’s move is part of an industry-wide inflection point. By mid-2026, all three major consumer AI labs have shipped persistent memory systems :

• Anthropic launched Claude Dreaming on May 6, 2026, using offline curation that reviews up to 100 past sessions to identify patterns and self-improve.
• Google introduced Gemini Spark’s account-level context at I/O on May 19, 2026, embedding memory directly into its assistant ecosystem.
• OpenAI now matches and, in some respects, leapfrogs both with a system that synthesizes narrative profiles in the background for hundreds of millions of users.

This is not a niche feature race. Persistent memory represents a fundamental shift from stateless to stateful AI—from tools that answer questions to assistants that build an ongoing relationship with each user .

Strategic significance: the personalized assistant vision

Dreaming V3 is one of the clearest signals yet that OpenAI is building toward a proactive, context-aware personal assistant rather than a general-purpose chatbot . By reducing reliance on explicit commands and instead inferring context automatically, the system approaches a long-standing vision: an AI that knows your work projects, your goals, your dietary restrictions, your travel plans, and your communication style—without being retold each session.

The combination of dramatically lower compute costs and narrative-level understanding makes this vision economically viable at global scale. For OpenAI, personalized memory isn’t just a feature—it’s the foundation for agent-like behavior, where an AI can act on your behalf with genuine understanding of who you are and what you need .

Access and rollout