MemAgent: Cognitive Models for Human-Like Decision-Making (IT:U Seed) 

Human decisions are shaped by memory processes – what we know, what we forget, and how we retrieve information in context. Still, most AI systems treat human cognition as static or overly simplified, ignoring the nuanced ways people recall knowledge and apply rules in dynamic environments. While recent advances in generative agents can partly simulate human behavior via large language models, they often rely on abstract or artificial memory representations that fail to reflect real cognitive processes. 

MemAgent aims to bridge this gap by modeling human memory more realistically, based on theories from cognitive science to simulate how people store, retrieve, and use information during decision-making. The project combines generative agents with psychologically grounded memory architectures to create more authentic simulations of human rule-based and knowledge-driven behavior across diverse task environments. 

These agents will not only model how humans behave but also research how individual goals, prior experiences, and knowledge drive context-specific memory activation. Based on this foundation, MemAgent will develop two applied outcomes: (1) context-aware notification systems that base decisions on conversation histories and personalized relationship models and (2) personalized digital twins that assist in decision-making, even when the user is not present. 

By advancing the fidelity of human-like agents, MemAgent offers a powerful framework for understanding cognition and enhancing real-world applications, from smarter assistive systems to ethically aligned AI collaborators. Ultimately, this project explores how modeling memory isn’t just about simulating thought, it’s about building technology that truly understands how we think.