OpenAI's old memory system worked like a notebook you had to write in yourself. Dreaming V3 works like a roommate who reads your journal while you sleep, learns everything about you, and promises to only use it to be more helpful. Totally normal. Very fine.
The Background Process That Never Forgets Anything You Ever Said
OpenAI began rolling out Dreaming V3 to ChatGPT Plus and Pro subscribers on June 4, 2026. Unlike the previous saved-memories system — which required users to explicitly instruct ChatGPT to remember something — Dreaming V3 runs as an asynchronous background process that continuously analyzes a user's entire conversation history and synthesizes a unified memory state, updated automatically over time.
The system handles temporal context intelligently: if a user once told ChatGPT they were "going to Singapore in July," Dreaming V3 automatically converts that entry to "went to Singapore in July 2026" after the date passes. Performance gains are dramatic — factual recall jumped from 41.5% to 82.8%, preference adherence from 31.4% to 71.3%, and time-sensitive context accuracy from 9.4% all the way to 75.1%, all at 5x lower compute cost than the previous architecture.
Power Features, and One Very Unsettling Privacy Footnote
The synthesized memory profile is injected into the system prompt at the start of every new conversation, meaning ChatGPT arrives already knowing your habits, projects, travel history, and preferences without a single prompt from you. New controls include a Memory Summary page, Gmail inbox integration for Plus and Pro users in supported regions, and source-level transparency for what the system recalls. That sounds reasonable until you read the fine print: OpenAI has acknowledged that the Memory Summary page "may not include everything ChatGPT remembers" — a curious qualifier for a privacy control.
Security researchers have flagged a prompt injection risk: malicious content encountered via third-party sources — a shared document, a linked webpage — could potentially update the persistent memory state across future sessions. Separately, disabling saved memories does not disable the "Improve the model" training toggle, which requires an independent opt-out. Free-tier and international users receive the rollout by late July 2026.
Dreaming V3 is a genuine leap in conversational AI personalization. Just read the settings page carefully before it decides that terse 2024 exchange with a customer service bot is foundational personality context forever.
Source: OpenTools