Memp introduces a novel approach to equip AI agents with procedural memory, drawing inspiration from human cognition to enable adaptation to new tasks and environments. This framework addresses the limitations of current Large Language Model (LLM) agents, which often struggle with rigid, manually designed memories, leading to inefficiency and fragility. Memp tackles this by allowing agents to learn from past experiences, distilling them into detailed instructions and abstract scripts.
The Memp framework focuses on building, retrieving, and dynamically updating procedural memory. This involves transforming past actions into step-by-step instructions and higher-level strategies, which are then stored and refined over time. The system continuously updates, corrects, and discards information, ensuring the memory repository evolves with new experiences. Evaluations on tasks like TravelPlanner and ALFWorld demonstrate that agents using Memp achieve higher success rates and efficiency as their memory improves.
Furthermore, procedural memory built on a strong model can be transferred to a weaker model, yielding significant performance gains. This highlights the potential for creating more adaptable, efficient, and human-like AI agents. By optimising memory construction, retrieval and updating, Memp allows agents to refine and reuse past experiences, improving accuracy in long-horizon tasks.
