Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution
Summary: arXiv:2512.10696v2 Announce Type: replace
Abstract
Procedural memory empowers large language model (LLM) agents to internalize valuable “how-to” knowledge, theoretically minimizing redundant trial-and-error processes. However, current frameworks predominantly operate under a “passive accumulation” paradigm, viewing memory as a static append-only archive. To address the limitations of static storage and enhance dynamic reasoning, we introduce ReMe (Remember Me, Refine Me), a holistic framework designed for experience-driven agent evolution.
ReMe Framework Overview
The ReMe framework innovates across the memory lifecycle through three key mechanisms:
- Multi-faceted Distillation: This mechanism extracts detailed experiences by identifying success patterns, analyzing failure triggers, and generating comparative insights that inform future actions.
- Context-Adaptive Reuse: ReMe tailors historical insights to new contexts through scenario-aware indexing, ensuring that past experiences are relevant and actionable in varying situations.
- Utility-Based Refinement: This mechanism autonomously adds valid memories and prunes outdated ones, maintaining a compact and high-quality pool of experiences essential for effective decision-making.
Experimental Validation
Extensive experiments conducted on the BFCL-V3 and AppWorld platforms demonstrate that the ReMe framework establishes a new state-of-the-art in agent memory systems. One of the most significant findings from our research is the memory-scaling effect observed during testing. Agents equipped with ReMe, such as Qwen3-8B, outperformed larger, memoryless counterparts like Qwen3-14B.
Significance of Memory-Sampling
This outcome suggests that self-evolving memory mechanisms can provide a computationally efficient pathway for lifelong learning, allowing agents to adapt and refine their knowledge base over time. The ability to evolve and improve memory systems is crucial for developing more capable and intelligent agents that can operate effectively in dynamic environments.
Open Source Initiative
In the spirit of collaboration and advancement in the field of artificial intelligence, we are pleased to announce the release of our code and the reme.library dataset. This initiative aims to facilitate further research and foster innovation within the AI community.
Conclusion
The ReMe framework represents a significant step forward in the evolution of procedural memory for LLM agents. By transitioning from passive memory storage to dynamic reasoning and experience-driven evolution, ReMe offers a robust solution that enhances the effectiveness of AI agents in real-world applications.