Multi-Agent EMPO²: Collaborative Exploration Through Shared and Private Memory

TL;DR: Imagine a team of EMPO² agents sharing hints but also keeping secrets to solve tasks faster! We’ll explore letting agents share parts of their memory while retaining private experiences, hypothesizing improved collective exploration and robustness. Initial tests could involve multi-agent ScienceWorld scenarios, measuring group performance and knowledge transfer.

Research Question: Can hybrid shared-private memory architectures among EMPO² agents boost collective exploration and generalization across diverse tasks compared to fully isolated or fully shared memory settings?

Hypothesis: Selectively sharing memory representations between agents—while retaining private, individualized experiences—will enhance collective exploration efficiency and adaptability, especially in tasks requiring diverse strategies or knowledge transfer.

Experiment Plan: - Design a multi-agent framework where agents have both private and shared memory modules.

• Develop strategies for selective sharing (e.g., only sharing high-utility or novel experiences).
• Test in multi-agent environments (e.g., collaborative ScienceWorld tasks, WebShop with multiple agents).
• Analyze team task completion rate, diversity of explored states, and knowledge transfer efficiency.

References:

• Liu, Z., Kim, J., Luo, X., Li, D., & Yang, Y. (2026). Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization.
• Hu, S., Fang, Z., Fang, Z., Deng, Y., Chen, X., & Fang, Y. (2024). AgentsCoDriver: Large Language Model Empowered Collaborative Driving with Lifelong Learning. arXiv.org.