KEEP: A KV-CACHE-CENTRIC MEMORY SYSTEM FOR EMBODIED PLANNING

KEEP is a high-performance memory management system designed to enable efficient long-horizon task planning for LLM-powered embodied agents. By introducing a KV-cache-centric memory architecture, KEEP significantly reduces inference latency while maintaining planning accuracy in dynamic environments.

Key Features 🚀

• Static-Dynamic Memory Construction: Groups memory by update frequency to minimize KV cache invalidation

• Multi-Hop Memory Recomputation: Dynamically retrieves and recomputes critical memory interactions through iterative importance propagation

• Layer-Balanced Pipeline Scheduling: Eliminates computation bubbles via cross-layer prefetching and balanced loading

Why KEEP? 📚

While LLMs show great promise for embodied planning, traditional context-window memory approaches lead to long prompts and high latency. KEEP directly addresses this by:

• Enabling >2× faster planning than baseline methods (e.g., on ALFRED benchmark)

• Maintaining competitive task success rates with significantly lower TTFT

• Providing scalable KV memory management for long-horizon tasks

Install

KEEP_transformers is a simple implementation of KEEP based on Transformers for accuracy evaluation.

KEEP_vllm is a implementation on vLLM for latency evaluation.

The two implementations require different environments. Please follow the setup instructions inside each respective file.

Citation

If you find this work useful for your research, please cite our paper:

@article{yang2025keep,
  title={KEEP: A KV-Cache-Centric Memory Management System for Efficient Embodied Planning},
  author={Yang, Zebin and Xie, Tong and Lu, Baotong and Liu, Shaoshan and Yu, Bo and Li, Meng},
  journal={arXiv preprint arXiv:2602.23592},
  year={2025}
}