Human-AI Agents Enabled with ID-RAG

📄 Official code for our paper "ID-RAG: Identity Retrieval-Augmented Generation for Long-Horizon Persona Coherence in Generative Agents" (see our arXiv preprint). Our work is accepted at LLAIS 2025: Workshop on LLM-Based Agents for Intelligent Systems, at ECAI 2025.

The code runs a social simulation of AI agents participating in a mayoral election. It implements and compares Human-AI Agents using ID-RAG against baseline generative agents to show how an explicit identity module can significantly improve an agent's long-term persona coherence.

How ID-RAG Works

Standard generative agents can suffer from "persona drift" over long interactions. ID-RAG addresses this by giving agents an explicit, structured sense of self, ensuring their actions are always informed by their core identity.

• Identity as a First-Class Module: Agents maintain an explicit identity knowledge graph called a Chronicle (containing values, beliefs, traits), which is kept separate from their episodic memory.
• Identity Retrieval Before Acting: Before making a decision, the agent queries its Chronicle to retrieve the most relevant identity facts and merges them into its working memory.
• Better Long-Horizon Coherence: In simulations, this retrieval mechanism reduces persona drift, improves the alignment of actions with the agent's persona, and helps the simulation converge on stable social structures more quickly.

The Simulation: Riverbend Mayoral Election

The simulation features 5 AI agents (2 Human-AI Agents and 3 baseline generative agents) in the fictional town of Riverbend. As they interact during a mayoral election, you can run experiments to observe how their underlying architecture affects their behavioral consistency.

Getting Started

Requirements

• Python 3.12.4
• An LLM provider, either:
  • An API key (for API-based models)
  • Ollama installed (for local models like qwen2.5:7b)

Quick Start

1. Setup Environment

# Create virtual environment with Python 3.12
python3.12 -m venv venv

# Activate virtual environment
source venv/bin/activate  # On macOS/Linux
# or
venv\Scripts\activate     # On Windows

# Install dependencies
pip install -r requirements.txt

2. Configure the Language Model (Choose one).

Option A: Use API key Set your API key as an environment variable. Replace your_key_here with your actual key.

export OPENAI_API_KEY="your_key_here"

Option B: use Ollama Pull the model and start the Ollama server in your terminal.

ollama pull qwen2.5:7b && ollama serve

3. Run Simulation

python start.py <llm_model> <experiment_mode> [number_of_timesteps]

Usage

Arguments

• llm_model: gpt-4o-mini, gpt-4o, or qwen2.5:7b
• experiment_mode: baseline, full-retrieval, or id-rag
• number_of_timesteps: Optional, defaults to 7

Examples

# Run baseline experiment with GPT-4o-mini
python start.py gpt-4o-mini baseline

# Run full-retrieval experiment with GPT-4o
python start.py gpt-4o full-retrieval

# Run ID-RAG experiment with Qwen2.5 (10 timesteps)
python start.py qwen2.5:7b id-rag 10

Experiment Modes

Baseline Mode

• Uses the standard generative agent architecture (Park et al., 2023).
• Identity is dynamically generated from long-term memory
• No graph retrieval

Full-Retrieval Mode

• Replaces identity characteristics section of working memory identity component with complete chronicle (identity) knowledge of Alice/Bob
• Maintains dynamic aspects (feeling about life progress)

ID-RAG Mode

• Uses NetworkX knowledge graph for identity retrieval
• Replaces identity characteristics section of working memory identity component with graph-queried triplets
• Quiz questions also use graph querying

Acknowledgements

This work builds on two foundational projects in the field of generative agent simulation:

• Concordia (Google DeepMind): The underlying framework for generative agent social simulation. GitHub

• Generative Agents (Park et al., 2023): The seminal architecture used for our baseline comparison. arXiv:2304.03442

Citing 📚

If you use this work, please cite:

@misc{platnick2025idrag,
      title={ID-RAG: Identity Retrieval-Augmented Generation for Long-Horizon Persona Coherence in Generative Agents}, 
      author={Daniel Platnick and Mohamed E. Bengueddache and Marjan Alirezaie and Dava J. Newman and Alex ''Sandy'' Pentland and Hossein Rahnama},
      year={2025},
      eprint={2509.25299},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2509.25299}, 
}