Reproduce “MemFactory: Unified Inference & Training Framework for Agent Memory” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ziliang Guo, Ziheng Li, Bo Tang, Feiyu Xiong, Zhiyu Li

The MemFactory framework provides a unified infrastructure for developing, training, and evaluating memory-augmented AI agents, particularly those leveraging Reinforcement Learning for adaptive memory management. Optimized agents demonstrated performance improvements, with Qwen3-1.7B seeing a 14.8% average score increase and Qwen3-4B-Instruct a 7.3% increase on long-context tasks.

Key insights

  • A highly modular and 'Lego-like' design significantly reduces the engineering burden in Memory-RL research by enabling flexible construction and customization of diverse memory architectures.
  • Group Relative Policy Optimization (GRPO) allows for computationally efficient RL training, making sophisticated memory policy optimization accessible even with limited resources, such as a single A800 GPU.
  • Unified frameworks can streamline the integration, training, and evaluation of complex memory pipelines, fostering systematic research and accelerating the discovery of more effective memory management strategies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.