Agent memory that forgets and distorts on purpose (github.com)

A groundbreaking approach in AI memory management was announced with the introduction of Selective Memory (SelMem), which allows large language model (LLM) entities to purposefully forget or distort past experiences. Unlike traditional memory systems that prioritize fact retention for maximum coverage and coherence, SelMem emphasizes a curated and subjective recollection of events. This technique enables different instances of the same model to diverge based on personal narratives, allowing for a "path-dependent" evolution of identity and context, which fosters unique interactions and responses. For the AI and machine learning community, this development is significant as it challenges conventional assumptions about memory and identity in AI systems. With features such as memory decay, mood-dependent recall, and selective history shaping, SelMem empowers models to produce responses that are not only contextually relevant but also imbued with personalized storytelling qualities. Technical details include the use of Rust for implementation with SQLite for data storage, creating a system where memory is dynamically managed in RAM while archives are sealed away, effectively simulating a more human-like memory experience where perception and memory shape identity.