Memory-augmented LLM Agents: Enabling Persistent Individuality in Simulated Behaviors

TL;DR: What if simulated users could “remember” their past actions, choices, and contexts over long time horizons, like real people? The experiment would augment LLM agents with persistent, structured memory modules to test whether this preserves unique behavioral trajectories and staves off persona homogenization.

Research Question: Can persistent, individualized memory modules—tracking each agent’s long-term behavioral history—help LLMs maintain consistent, non-average personas over long-horizon simulations?

Hypothesis: Augmenting LLMs with agent-specific memory structures will reinforce individualized decision patterns and reduce convergence toward an average persona, leading to higher fidelity in long-horizon, cross-scenario behavioral simulations.

Experiment Plan: - Implement a memory-augmented agent architecture where each simulated user is assigned a memory module (e.g., a structured database or context window) recording their actions, choices, and outcomes across scenarios.

• Simulate the OmniBehavior benchmark tasks using both vanilla LLM agents and memory-augmented agents.
• Quantitatively measure individuality (e.g., behavioral trajectory divergence, clustering, entropy) and compare against real-world users.
• Evaluate whether memory augmentation prevents homogenization and increases retention of unique behavioral signatures.

References:

• Chen, J., Xu, R., Cao, B., Pan, R., Zhang, Y., Hu, Y., Du, Y., Gao, T., Lu, Y., Sun, Y., Han, X., Sun, L., Wu, X., & Lin, H. (2026). Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces.
• Wang, C., Zeng, S., & Wang, C. (2024). Large Language Models Need Consultants for Reasoning: Becoming an Expert in a Complex Human System Through Behavior Simulation. arXiv.org.