😺 ChatGPT admitted its memory was broken
The Neuron · Grant Harvey · 2026-06-07
OpenAI publicly discloses that ChatGPT's memory feature had only 41.5% factual recall accuracy in 2024 and announces Dreaming V3, a background synthesis system that raises accuracy to 82.8% and extends memory access to free users.
Appears in
Extraction
Topics: chatgpt-memoryopenai-product-updatesai-personalizationai-safety-policybot-traffic
Claims
- ChatGPT's original memory feature had a factual recall accuracy of only 41.5% in 2024, meaning it was wrong more than half the time on memory-dependent tasks.
- OpenAI's new Dreaming V3 system automatically synthesizes conversation history in the background without user commands, raising factual recall to 82.8% by 2026.
- Preference adherence—how well ChatGPT follows a user's style and habits—improved from 55.3% to 71.3% with the new memory system.
- Compute costs for the memory system dropped 5x, enabling free users to access memory features for the first time.
- Anthropic called for a global pause on frontier AI development, warning that models are approaching the ability to self-improve without human oversight.
- Bots now generate 51% of all internet traffic, marking the first time humans are a minority online.
Key quotes
OpenAI just published data showing its flagship feature was wrong more than half the time in 2024. They fixed it. But every AI assistant is running some version of this problem right now, quietly, with no warning light.
Memory plus financial data is a significant combination; an AI that knows your work, your habits, and your money looks a lot less like a chatbot.
Bots now generate 51% of all internet traffic, the first time humans are a minority online; AI shopping agents are hitting thousands of retail sites per second, quietly making your analytics dashboards unreliable.