AI agents with human-like memory set to transform city transport planning

The research proposes a new cognitive architecture that harnesses large language models to simulate more human-like decision-making. Unlike traditional agents, these LLM-powered agents use a layered memory system consisting of short-term episodic memory and long-term semantic memory. This allows them to store daily experiences, reflect on past choices, and adapt future decisions accordingly.


CO-EDP, VisionRICO-EDP, VisionRI | Updated: 16-09-2025 23:05 IST | Created: 16-09-2025 23:05 IST
AI agents with human-like memory set to transform city transport planning
Representative Image. Credit: ChatGPT

Urban mobility systems are under increasing strain as cities grapple with congestion, sustainability, and equity concerns. A new study argues that traditional tools for modeling transportation behavior are no longer sufficient. Instead, artificial intelligence can provide a breakthrough. Researchers from the Valencian Research Institute for Artificial Intelligence at the Universitat Politècnica de València and VALGRAI present a novel approach that integrates large language models into mobility simulations.

The paper, titled “Cognitive Agents in Urban Mobility: Integrating LLM Reasoning into Multi-Agent Simulations,” was published in Sensors. It introduces a framework that enables virtual agents to reason, adapt, and plan like real commuters. By embedding LLM reasoning and multi-horizon memory into the SimFleet simulator, the researchers demonstrate that urban mobility models can better capture how people respond to disruptions and adjust routines over time.

Why do traditional mobility models fall short?

Conventional agent-based models (ABMs) have long been used to test urban transport systems. These models simulate commuters, vehicles, and infrastructure, but they often rely on pre-defined rules. Agents typically act in predictable ways, repeating the same routes and failing to reflect the nuanced decision-making that humans demonstrate in real cities.

This rigidity becomes a critical weakness when disruptions occur. For example, a taxi strike or road closure in real life leads to diverse responses, some commuters reschedule, others change transport modes, and a few simply cancel trips. ABMs, however, tend to model such shocks with unrealistic uniformity.

The authors argue that more intelligent, adaptable agents are essential for evaluating how policies and crises affect urban transport. Without capturing adaptive behavior, simulations risk providing misleading insights for city planners tasked with designing resilient and equitable systems.

How do LLMs transform cognitive agent design?

The research proposes a new cognitive architecture that harnesses large language models to simulate more human-like decision-making. Unlike traditional agents, these LLM-powered agents use a layered memory system consisting of short-term episodic memory and long-term semantic memory. This allows them to store daily experiences, reflect on past choices, and adapt future decisions accordingly.

To test this architecture, the researchers designed a population of 320 agents representing diverse sociodemographic groups, including students, elderly residents, administrative staff, and factory workers. Over a simulated 20-day period, these agents developed travel routines, made decisions based on memory, and adapted when faced with disruptions.

In one scenario, an 80 percent taxi strike forced agents to reconsider their mobility options. Agents equipped with cognitive memory adapted quickly, altering routes, changing departure times, or switching to other transport modes. By contrast, static or short-memory agents showed limited flexibility and struggled with punctuality.

The findings underscore the importance of memory-driven reasoning. Agents that could recall and reflect on past disruptions achieved greater efficiency and reliability in meeting their mobility goals. This, the study notes, brings simulated commuters far closer to the adaptive behaviors observed in real urban populations.

What are the implications for urban mobility planning?

The study calls LLM-driven agents a significant advance in the field of transport simulation. By incorporating reflection, adaptation, and identity-based diversity, the framework provides both greater realism and more transparent interpretability.

This has direct implications for policymakers and planners. Cities need tools that can evaluate resilience under stress conditions, whether caused by strikes, extreme weather, or infrastructure failures. Simulations that include adaptable, heterogeneous agents are better suited to testing strategies for equity, sustainability, and efficiency.

The integration of sociodemographic diversity is another key contribution. By modeling groups with distinct roles and constraints, such as students with class schedules or elderly residents with limited mobility, the framework highlights differences in vulnerability and resilience. This helps decision-makers identify which communities bear the brunt of disruptions and how interventions can be targeted to reduce inequalities.

The authors suggest that the approach could also be extended beyond mobility to other domains where agent-based simulations are used, such as emergency response or energy systems. The underlying principle remains the same: AI-driven agents with memory and reasoning can better capture the complexity of human decision-making under uncertainty.

  • FIRST PUBLISHED IN:
  • Devdiscourse
Give Feedback