KEEP: A KV-Cache-Centric Memory Management System for Efficient Embodied Planning

ArXi:2602.23592v2 Announce Type: replace-cross Memory-augmented Large Language Models (LLMs) have nstrated remarkable capability for complex and long-horizon embodied planning. By keeping track of past experiences and environmental states, memory enables LLMs to maintain a global view, thereby avoiding repetitive exploration. However, existing approaches often the memory as raw text, leading to excessively long prompts and high prefill latency. While it is possible to and reuse the KV caches, the efficiency benefits are greatly undermined due to frequent KV cache updates.