ChatGPT’s New Dreaming Memory Boosts Factual Accuracy to 83%

OpenAI officially launched a new memory architecture called Dreaming, which evolves the ChatGPT memory feature from the manually saved "Saved Memories" introduced in April 2024 to an automated, brain‑like system in Dreaming V3.

Saved Memories required users to issue explicit commands such as “remember I’m going to Singapore in July”; the system stored the phrase only for the duration of the conversation and ignored anything not explicitly saved, leading to stale notes.

In April 2025 OpenAI released Dreaming V0, shifting the approach from manual prompts to a background process that periodically scans chat history, extracts salient facts (e.g., a user’s love for wildlife photography, preference for quiet dinner settings, or need for strong air‑conditioning), and synthesizes them into a coherent memory state without user intervention.

Dreaming V0 improved personalization but could not operate as a full‑scale memory system because of high computational cost and limited coverage.

Dreaming V3, the latest release, delivers roughly five‑times higher computational efficiency (an 80% reduction in compute) and raises the memory quality across three dimensions defined by OpenAI: continuity of context, preference adherence, and timeliness of updates.

Measured accuracies for Dreaming V3 are 82.8% for context continuity, 71.3% for following user preferences, and 75.1% for keeping information up‑to‑date. The system now serves all users, including free accounts, after the recent optimization.

Concrete examples illustrate the improvements: after a user previously discussed underwater photography gear, asking “what TTL accessories do I need?” yields specific trigger and converter recommendations; a travel plan that previously required manual re‑specification now automatically incorporates a user’s stated preferences for wildlife photography, strong hotel AC, and quiet restaurants.

The Memory Summary page lets users view, edit, or highlight key memory items and directly query the model for deeper details.

Traditional memory suffered from static timestamps—e.g., a birthday party scheduled for “next Saturday” remained unchanged after the date passed—whereas Dreaming V3 updates temporal information automatically, correcting “you will go to Singapore next July” to “you went to Singapore in July 2026” and providing timely recommendations such as finding a still‑open food delivery service based on the user’s current location and time zone.

Overall, Dreaming V3 represents OpenAI’s most powerful memory system to date, combining higher accuracy, lower cost, and broader availability.