Cognitive Science-Inspired Memory Replay: Integrating Ebbinghaus Forgetting Curves into Self-Navigating Agents

Research Question: Does integrating memory decay mechanisms inspired by the Ebbinghaus forgetting curve into AgentEvolver’s experience reuse enhance exploration efficiency and long-term performance?

Hypothesis: Agents that prioritize replay of experiences according to a biologically-inspired forgetting curve will exhibit better retention of critical behaviors and improved sample efficiency compared to those using uniform or recency-based experience replay.

Experiment Plan: - Modify the self-navigating mechanism of AgentEvolver to weight past experiences for replay according to an Ebbinghaus-inspired forgetting function.

• Use benchmarks from AgentGym and compare the rate of re-learning previously encountered but rare tasks, overall exploration efficiency, and adaptability to environment changes.
• Analyze if the agent’s memory management leads to faster recovery from catastrophic forgetting or loss of rare skills.

References:

• Liang, X., Tao, M., Xia, Y., Shi, T., Wang, J., & Yang, J. (2024). Self-evolving Agents with reflective and memory-augmented abilities. arXiv.org.
• Zhai, Y. et al. (2025). AgentEvolver: Towards Efficient Self-Evolving Agent System.