Adaptive Memory Pruning for Continual Learning in EMPO² Agents

TL;DR: What if agents could "forget" just the right things to stay adaptable and efficient? Let's design EMPO² agents that automatically prune their memory, keeping only the most relevant experiences for future tasks. The initial experiment would compare static vs. adaptive memory management policies on continual learning benchmarks, hypothesizing that adaptive pruning yields better generalization and lower memory overhead.

Research Question: How can EMPO² agents dynamically manage and prune their episodic memory to maximize adaptability and performance in continual learning scenarios without accumulating irrelevant or outdated experiences?

Hypothesis: Adaptive memory pruning—guided by relevance scores or utility predictions—enables EMPO² agents to maintain high out-of-distribution performance and sample efficiency, outperforming both static and naive memory management strategies.

Experiment Plan: - Implement adaptive memory pruning strategies (e.g., relevance scoring, novelty detection, or task-specific utility).

• Test on environments where task distributions shift over time (e.g., continual task suites or evolving ScienceWorld tasks).
• Measure adaptability (success rate on new tasks), memory footprint, and sample efficiency, comparing static, random, and adaptive pruning.
• Expect adaptive pruning to enable efficient memory usage and improved performance on new tasks.

References:

• Liu, Z., Kim, J., Luo, X., Li, D., & Yang, Y. (2026). Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization.
• Zhang, D., Chen, L., Zhang, S., Xu, H., Zhao, Z., & Yu, K. (2023). Large Language Models Are Semi-Parametric Reinforcement Learning Agents. Neural Information Processing Systems.
• Zhang, M., Qian, F., & Liu, Q. (2024). Memory Sequence Length of Data Sampling Impacts the Adaptation of Meta-Reinforcement Learning Agents. arXiv.org.