Show HN: Unified multimodal memory framework, without embeddings (github.com)

MemU is a newly introduced agentic memory framework designed to enhance LLM and AI agent backends by processing and organizing multimodal inputs—including conversations, documents, images, audio, and video—into a structured hierarchical memory system. This framework features a three-layer architecture that enables flexible organization and traceability of extracted memory, supporting both embedding-based retrieval (RAG) for speed and non-embedding retrieval (LLM) for deep semantic understanding. The seamless integration of these two methods allows users to effectively extract and retrieve insights across various data types, making MemU a versatile tool for AI applications. The significance of MemU lies in its ability to unify diverse modalities into a singular memory structure, enhancing the efficiency and effectiveness of AI-driven tasks such as personal assistant interactions, customer support, and knowledge management. With an impressive 92.09% accuracy on the Locomo benchmark, MemU not only promises organizational benefits but also offers cash rewards and recognition through its upcoming New Year Challenge, encouraging community involvement and contributions to further its development. This collaborative approach could foster innovation and practical applications within the AI/ML landscape, showcasing the potential of advanced memory architectures in AI systems.