Beyond RAG and Vector-based memories for Active Learning

Large Language Models demonstrate remarkable capabilities in language understanding and reasoning, yet their knowledge remains largely static after training. Current solutions, such as Retrieval-Augmented Generation and vector-based memory systems, extend model capabilities but face fundamental limitations in scalability, knowledge organization, and adaptive learning. Inspired by principles of human cognition, this project proposes to investigate novel memory architectures for LLMs that move beyond factual storage toward structured, contextual, and abstract representations that support reasoning, planning, and continual adaptation. The research will explore how such memory mechanisms can drive active learning and multi-hop reasoning, enabling models to autonomously identify knowledge gaps, acquire new information, and consolidate it over time. By advancing memory-centric foundations for trustworthy and adaptive AI systems, the project aims to contribute high-impact scientific knowledge and open new directions in lifelong and human-aligned artificial intelligence.