LatentMem-Qwen3-4B

This is the official repository for LatentMem-Qwen3-4B, based on the Qwen3-4B backbone and optimized using Latent Memory Policy Optimization (LMPO).

LatentMem is a learnable multi-agent memory framework designed to customize agent-specific memories in a token-efficient manner. It comprises an experience bank that stores raw interaction trajectories and a memory composer that synthesizes compact latent memories conditioned on retrieved experience and agent-specific contexts.

Resources

• Paper: LatentMem: Customizing Latent Memory for Multi-Agent Systems
• Code: Official GitHub Repository

📂 Repository Structure

The repository is organized into two primary directories:

• data/: Contains the Raw Trajectories collected during the initial data collection phase. These include state-action pairs, environmental feedback, and intermediate reasoning steps used for training.
• model/: Contains the LatentMem Weights. These are the final weights after undergoing LMPO.

Citation

If you find this work helpful, please consider citing the paper:

@misc{fu2026latentmemcustomizinglatentmemory,
      title={LatentMem: Customizing Latent Memory for Multi-Agent Systems}, 
      author={Muxin Fu and Guibin Zhang and Xiangyuan Xue and Yafu Li and Zefeng He and Siyuan Huang and Xiaoye Qu and Yu Cheng and Yang Yang},
      year={2026},
      eprint={2602.03036},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.03036}, 
}