Latest Review on VLM Memory Mechanisms: Understanding Memory Through the Four Dimensions of When, Where, What, and How
14 hour ago / Read about 0 minute
Author:小编   

Researchers from eight institutions have jointly released a systematic review of memory mechanisms in vision-language models (VLMs). The review proposes a new 4D classification framework that systematically organizes over 300 related works along four dimensions—memory retention duration, storage location, stored content, and full-lifecycle operations. It summarizes eight typical forms of memory and compiles nearly 70 evaluation benchmarks, covering eight prominent application scenarios such as long video understanding, multimodal RAG, and embodied intelligence. The review notes that current VLM memory mechanisms face challenges including the tendency to omit key information, difficulty in making information trade-offs, hard-to-detect false memories, and single-dimensional existing evaluations. It further proposes future research directions such as building human-like hierarchical memory structures and improving the traceability and editability of memory.