ChatGPT Now Builds a Narrative Profile of You — Here's What It Knows

ChatGPT is no longer just remembering isolated facts about you. Through a background process called “Dreaming,” it is now assembling coherent prose profiles organized by life domain — and the accuracy jump in its personal context retrieval is substantial.

What Happened

OpenAI rolled out a significant update to ChatGPT’s memory system for Plus and Pro subscribers in the US, with free users to follow. The change is architectural: instead of storing bullet-point facts about users, the system now generates narrative profiles derived from conversation history. These profiles are organized into categories — work, hobbies, travel, education — and updated automatically in the background without requiring explicit user prompts.

The performance numbers behind the change are notable. Fact retrieval accuracy improved from 41.5% to 82.8% year-over-year. The system’s ability to incorporate personal preferences jumped from 31.4% to 71.3%, and memory freshness improved from 52.2% to 75.1%. OpenAI also reduced the required computing by a factor of five to enable the broader rollout.

Users can access a new summary page showing exactly what ChatGPT has stored about them, with controls to correct specific memories, opt out of individual entries, or toggle chat history and memory separately under Personalization settings.

Why It Matters

This is the memory architecture that makes proactive AI viable. The accuracy gap that existed in earlier memory systems — where the model would forget or misremember user context even within short sessions — is the same gap that has kept AI systems from operating reliably as persistent background agents. The jump from 41.5% to 82.8% in fact retrieval is not a quality improvement; it is a threshold crossing.

For users, the immediate implication is that ChatGPT’s personalization now operates more like a well-briefed assistant than a tool that periodically forgets your name. For developers and enterprise teams, the more significant implication is that the underlying memory infrastructure — narrative profiles, background updating, domain-organized retrieval — is the model for how persistent agent context will be managed at scale.

The privacy surface is also worth naming directly. A system that builds a continuously updated narrative profile from conversation history, covering work, personal interests, and travel preferences, holds more coherent information about a user than a list of remembered facts does. OpenAI has provided controls for correction and opt-out, but the default behavior is accumulation. Users who have previously treated ChatGPT as a stateless tool should revisit those assumptions.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.