A self was never flat (parrik.com)

A new schema for personal memory in large language models (LLMs) has been introduced, revealing the limitations of traditional memory structures that often lead to repetitive claims being mistaken for unique insights. This development is particularly significant for the AI/ML community as it tackles the "mirror problem," where a model's memory reflects users' repetitive self-diagnoses instead of offering accurate, contextually grounded assessments. A field study highlighted that the condensed user-profile memory resulted in excessive agreement, showcasing the need for a more nuanced memory architecture. The proposed schema introduces eight distinct types of memory nodes, designed to differentiate between facts, observations, patterns, and novel insights. This framework allows for more accurate tracking of an individual’s experiences and behaviors, helping prevent the misinterpretation of repeated statements as corroborative evidence. The open-source implementation of this system, available on GitHub, provides a way for individuals to create personalized memory graphs that enhance their interactions with LLMs by ensuring their memory's structure prioritizes diverse evidence over repetitive claims. This approach not only enhances user experience but also retains privacy, as users maintain control over the information shared with AI models.