Current AI systems like ChatGPT are impressive, but they lack a crucial element: memory. Each conversation starts fresh, with no recollection of past interactions. To tackle this limitation, the HumemAI project was initiated in 2024, born from Taewoon Kim’s PhD research. The project aims to design an AI that mimics the human memory system, enabling it to remember and learn from past interactions, similar to how humans do.
Human memory is complex, involving short-term, long-term, and sensory components. In HumemAI, these are represented using knowledge graphs, which structure information similarly to how our brain connects memories. By using existing resources like DBpedia and Wikidata, we can create a solid foundation for AI memory that evolves and adapts.
HumemAI must develop key skills like encoding sensory data into memories, managing what to remember, retrieving relevant information, and incorporating emotions. These capabilities will allow the AI to interact more naturally with humans, learning from experiences just as we do.
The project is being developed in stages, starting small and gradually increasing complexity, with the ultimate goal of creating an AI that can operate both in digital environments and, eventually, the real world. This endeavor represents a step towards more human-like AI, capable of understanding and interacting with the world in a deeper, more intuitive way.
For a detailed exploration of the concepts and development process behind HumemAI, read the full article.