In the dataset, 96% of memories were created unilaterally by the conversational system rather than at a user's explicit request. The authors classified 28% of entries as containing GDPR-defined personal data and 52% as containing psychological insights about participants. They found that 84% of memories were directly grounded in user context.
The authors propose Attribution Shield, a framework intended to anticipate sensitive inferences, alert users, and suggest alternative query wording. The paper does not establish that its sample represents all ChatGPT users, that every inferred memory is inaccurate or harmful, or how OpenAI operationally uses memory entries across its product.