AtomMem-8B

Model Overview

AtomMem-8B is specifically designed to handle complex, multi-step information processing by integrating the AtomMem framework—a learnable memory management system that deconstructs high-level memory workflows into fundamental atomic operations: Create, Read, Update, and Delete (CRUD).

Base Model: Built upon Qwen3-8B, inheriting its advanced linguistic capabilities and efficient architecture.

Agent Architecture: Implements the AtomMem paradigm, which reframes memory management as a dynamic decision-making problem.

Domain: Optimized for Multi-hop Question Answering (QA), specifically targeting the challenges of the HotpotQA dataset.

Training Paradigm: This model has undergone a two-stage training process:

Supervised Fine-Tuning (SFT): To align the model with atomic memory operation sequences.

Reinforcement Learning (RL): Optimized using RL (GRPO) on HotpotQA trajectories, enabling the agent to learn task-aligned memory policies that maximize reasoning accuracy and evidence retrieval efficiency.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{huo2026atommemlearnabledynamic,
      title={AtomMem : Learnable Dynamic Agentic Memory with Atomic Memory Operation}, 
      author={Yupeng Huo and Yaxi Lu and Zhong Zhang and Haotian Chen and Yankai Lin},
      year={2026},
      eprint={2601.08323},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.08323}, 
}