A Brain-inspired Multi-memory Agentic Framework for Lifelong Learning in Physical Embodied Systems

PDF view of the paper entitled Robomemory: a multi -memory framework inspired by the brain for lifetime learning in physical systems, by Mingong Lei and 14 other author

PDF HTML (experimental) view

a summary:We offer Robomemory, which is a multi -memory framework inspired by the brain to lifelong learning in physical physical systems, and deals with important challenges in realistic environments: continuous learning, multi -unit memory chronic, capturing tasks, and reducing an endless loop in closed planning. It is based on cognitive neuroscience, combining four basic units: the pre -processing processor (such as mulch), lifelong memory system (hippocampus), and a long -term closed planning unit (frontal lobe). The lifelong memory system, which is essential to the framework, reduces the problems of speed in reasoning in complex memory frameworks through updates/retrieval parallel via spatial and temporal sub -components, transverse, and three. It includes a graph of dynamic knowledge (KG) and consistent architectural design to enhance the consistency of memory and expansion. The assessments on the embodiment of Bechnench show the basic performance of the open sources (QWEN2.5-VL-72B-ins) by 25 % in the success rate in the average success rate and transcending the modern situation (SOTA) (CLAUDE3.5-Sonnet) by 5 %, and the creation of a new Sota. Huritage studies confirm the health of the main components (critic, spatial memory, long -term memory), while publishing in the real world confirms its ability to learn lifelong with highly improved success rates through frequent tasks. Robomemory reduces high time challenges with expansion, as it acts as a basic reference to integrate multimedia memory systems into physical robots.