Neuro-Inspired Episodic-Spatial Memory for LLM Agents: A Cognitive Synthesis

TL;DR: Can we make EMPO² agents remember like the human brain—combining where and when things happened? By fusing neuroscience-inspired episodic-spatial memory structures with EMPO², we could enable richer, more context-aware exploration. First experiments would benchmark against vanilla EMPO² on tasks demanding spatial reasoning and temporal credit assignment.

Research Question: Does integrating neuro-inspired episodic and spatial memory architectures into EMPO² agents enhance their context-aware exploration and temporal reasoning abilities in complex, partially observable environments?

Hypothesis: Agents equipped with episodic-spatial memory will demonstrate superior performance in tasks requiring recall of spatiotemporal contexts, outperforming standard EMPO² and memory-less baselines.

Experiment Plan: - Implement a memory module inspired by hippocampal place and time cells (e.g., memory traces labeled with spatial and temporal information).

• Evaluate on grid-based, navigation, or partially observable environments (e.g., ScienceWorld with spatial-temporal puzzles).
• Measure task completion, sample efficiency, and context recall accuracy.
• Compare with standard EMPO² and other memory-augmented approaches.

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.