Stock Markets June 4, 2026 12:30 PM

OpenAI launches Dreaming V3 to improve ChatGPT memory for subscribed users

New memory synthesis architecture aims to boost recall, preference adherence and currency while cutting compute needs

By Marcus Reed

OpenAI has begun deploying Dreaming V3, an upgraded memory synthesis system for ChatGPT that curates chat history automatically to reduce staleness and improve accuracy. The update is available to ChatGPT Plus and Pro users in the United States and will be rolled out to additional countries and Free users in the coming weeks. Performance metrics cited by the company show notable gains in factual recall, preference adherence and currency, alongside a roughly fivefold reduction in compute required to serve the feature.

OpenAI launches Dreaming V3 to improve ChatGPT memory for subscribed users

Key Points

  • Dreaming V3 is being rolled out for ChatGPT Plus and Pro users in the United States and will expand to more countries and Free users in the coming weeks - impacts consumer AI access and user experience.
  • OpenAI reports substantial performance improvements: factual recall task success at 82.8%, preference adherence at 71.3%, and currency at 75.1% compared with lower rates in 2025 and 2024 - relevant to AI product quality and enterprise adoption.
  • The compute required to serve dreaming was reduced by approximately five times, enabling broader availability to Free users and lowering operational cost pressures for the company - relevant to cloud compute and service economics.

OpenAI has started rolling out a new memory synthesis system for ChatGPT, calling the upgrade Dreaming V3. The release, which began on Thursday for ChatGPT Plus and Pro subscribers in the United States, is intended to improve how the chatbot retains and uses user preferences and conversational context across sessions.

According to the company, Dreaming V3 works by automatically curating relevant details from prior chat history in the background. The design specifically targets problems the firm identified with previous memory capabilities - namely staleness and accuracy - by letting the system surface and use context without users having to explicitly save items.

Memory functionality for ChatGPT was first introduced in April 2024 as saved memories, a feature that required users to instruct the assistant to store certain details. In April 2025, OpenAI deployed the first generation of dreaming, which enabled the model to reference chat context beyond the saved memories list. The new Dreaming V3 is described as a markedly more capable and more compute-efficient architecture compared with earlier iterations.

OpenAI provided comparative performance figures for the upgraded system. Factual recall task success is reported to have improved to 82.8%, compared with 67.9% in 2025 and 41.5% in 2024. Preference adherence has risen to 71.3%, up from 55.3% in 2025 and 31.4% in 2024. The system's ability to remain current over time reached 75.1%, versus 52.2% in 2025 and 9.4% in 2024.

In addition to accuracy gains, the company says it cut the compute required to operate dreaming by about five times. That reduction is cited as a key factor in making the feature practical to offer to Free users as well as paid subscribers. OpenAI also noted that dreaming will serve as a shared memory foundation for all users and that it will continue to iterate on the capability.

The initial rollout is limited to Plus and Pro users in the United States, with an expansion to other countries and Free users planned over the coming weeks. The company framed the update as both a performance and an efficiency upgrade without providing additional technical details beyond the comparative metrics and the compute reduction estimate.


Deployment context

OpenAI is positioning Dreaming V3 as a background mechanism that curates chat history automatically rather than relying solely on explicit saved memories. The company reports measurable improvements in three core areas - factual recall, preference adherence and staying current - and a substantial reduction in compute cost to support broader availability.

Risks

  • The rollout timeline beyond initial U.S. Plus and Pro users is described only as occurring over the coming weeks - availability and timing in other markets remain uncertain, affecting user access in international markets.
  • While the company cites improved metrics, the article does not provide technical specifics or independent validation of the reported gains - customers and enterprise buyers may face uncertainty when assessing reliability for production use.
  • The reduced compute requirement is presented as an approximate fivefold cut; the lack of detailed benchmarking or disclosure of trade-offs leaves open questions about potential limitations or costs shifted elsewhere in the system - this could affect cloud providers and cost modeling for deployments.

More from Stock Markets

Keystone Acquisition Raises $287.5 Million in Unit IPO; Private Warrant Sale Closed Jun 4, 2026 Anthropic President Points to Soaring AI Training Costs as Rationale for IPO Jun 4, 2026 Broadcom, Ciena and a Wide Range of Stocks See Big Moves on Thursday Jun 4, 2026 U.S. Jobs Report and Wage Data Take Center Stage Friday; Multiple Market Indicators to Follow Jun 4, 2026 Qantas Holds Talks on Purchase of About 20 Wide-Body Jets from Airbus or Boeing Jun 4, 2026