OpenAI's ChatGPT Memory V3 now profiles users across all conversations, raises accuracy and privacy concerns

OpenAI has deployed Dreaming V3, a background memory synthesis system that builds comprehensive user profiles from chat history. The company reports factual task recall jumped from 41% in 2024 to 82% in 2026, while reducing compute costs by 5X.

How the memory system evolved

ChatGPT's memory capability has progressed through three distinct phases:

2024 (Original Memory): Simple fact storage with no context. Each saved memory was an isolated data point that quickly became stale and irrelevant to future conversations.

2025 (Dreaming V0): The model began referencing chat history in the background without explicit prompting, automatically curating memories from conversations.

2026 (Dreaming V3): The current system performs data synthesis, effectively composing user dossiers. According to OpenAI, it tracks "dense, multi-session, multilayered long-running projects."

Performance claims

OpenAI reports significant improvements across three metrics:

• Factual task recall: 41% (2024) → 82% (2026)
• Accuracy over time: 9% task rate (2024) → 75% (2026)
• Preference adherence: 31% (2024) → 71% (2026)
• Compute cost: Reduced by 5X

The efficiency gains make it economically viable for OpenAI to roll out the feature to all users, including free tier accounts, over the coming weeks. The feature is currently available to Plus and Pro subscribers.

Accuracy concerns in testing

Independent testing by ZDNET revealed the system stores and retrieves inaccurate information. When asked about experience with Kasa smart plugs, ChatGPT claimed the tester had "moved that monitoring setup into Home Assistant" — a platform the user reports never installing. The model also added specific device model numbers not present in original saved memories.

OpenAI responded to accuracy concerns by stating users are seeing "a new high-level memory summary, rather than a complete inventory of facts ChatGPT may remember." The company noted that contextual details like tech stacks may still be used in conversations even when not displayed in the memory interface.

Privacy and control limitations

Disabling memory through Settings > Personalization > Memory does not fully purge stored information:

• Existing memories remain accessible to the model
• Chat history persists in OpenAI's systems
• Dream-based consolidation stops, but raw data remains
• Individual chats must be manually deleted to remove information

OpenAI's FAQ reveals an additional constraint: "Turning off Memory/Personalization does not disable safety features that may use limited, safety-relevant context in rare, high-risk situations."

Technical architecture

Dreaming V3 operates by attaching timestamps to information fragments and tracking temporal context alongside user interactions. The system continuously revises its internal model of each user, distinguishing between current preferences and outdated information through time-based decay signals.

What this means

OpenAI has achieved a legitimate technical breakthrough in persistent context management, with factual recall nearly doubling while cutting costs by 80%. However, the 82% accuracy rate means roughly one in five facts retrieved from memory could be wrong. For users relying on ChatGPT for technical work or decision-making, this error rate combined with opaque storage policies creates meaningful risk. The inability to fully disable memory even when the feature is turned off represents a significant shift in how conversational AI systems handle user data. Organizations deploying ChatGPT should audit what information employees share, as the system now builds cumulative profiles that persist indefinitely unless manually purged chat by chat.