Technical Summary: EconAI – Dynamic Persona Evolution and Memory-Aware Agents in Evolving Economic Environments

1. Problem Statement

Existing agent-based modeling (ABM) frameworks for economic simulation face a critical dichotomy:

• Rule-based models (e.g., LEN, CATS) rely on hand-crafted heuristics that lack flexibility and require extensive expert calibration for new scenarios.
• Learning-based models (e.g., Deep RL agents) offer adaptability but suffer from prohibitive computational costs and data requirements at the macro scale.
• Current LLM-driven approaches improve reasoning capabilities but fail to capture the nuanced interplay between short-term optimization and long-term strategic planning. Specifically, they often neglect the integration of event memory (historical experiences) and economic beliefs (confidence and sentiment), leading to simulation errors that accumulate over time and render models unrealistic.

The core challenge addressed is how to develop a unified framework that accurately simulates both macroeconomic dynamics and micro-level agent interactions while maintaining long-term coherence and adaptive behavior.

2. Methodology: The EconAI Framework

EconAI is a novel, LLM-powered simulation system that unifies macroeconomic and microeconomic environments. It models agents (households and firms) with a cognitive backbone comprising three core modules:

A. Event Perception and Memory Architecture

The system utilizes a dual-memory bank structure to manage historical context and immediate conditions:

• Long-term Memory Bank ($M_L$): Stores vector representations of high-level event summaries ($\phi(t_j, o_j)$) derived from historical interactions. These are encoded using text encoders (e.g., MiniLM) and retrieved via embedding-based mechanisms.
• Short-term Memory Bank ($M_S$): Retains contextual information for ongoing activities within a specific session.
• Event Summary Module: Unlike zero-shot LLM approaches, this module employs instruction tuning to generate structured summaries of economic activities, including task backgrounds, relevant activities, and specific summarization instructions.

B. Dynamic Persona and Economic Sentiment

To bridge the gap between objective data and subjective decision-making, EconAI introduces:

• Dynamic Persona Extraction: A module that models evolving agent preferences and economic beliefs, stored in a long-term persona bank. This allows for personalized, context-sensitive decision-making.
• Economic Sentiment Index (ESI): A quantitative metric ($ESIt$) that captures an agent's economic belief, updated via a decay factor ($\lambda$) and the LLM's latest sentiment assessment:
  $$ESIt = \lambda ESIt-1 + (1-\lambda)ESIt,LLM$$
  This index modulates decision-making:
  • Low Belief ($ESIt < 0$): Agents reduce consumption and increase work willingness.
  • High Belief ($ESIt > 0$): Agents increase consumption and reduce work.
  • Confidence Adjustment: Final decisions ($p_{wi}, p_{ci}$) are adjusted based on confidence levels and the ESI to prevent overreaction to short-term shocks.

C. Economic Environment and Decision-Making

The simulation operates within a unified framework involving households, firms, government, and financial institutions:

• Agent Decisions: Modeled as labor participation probability ($p_{wt}$) and consumption-to-income ratio ($p_{ct}$), generated by the LLM based on income, prices, savings, unemployment rates, interest rates, and the ESI.
• Macroeconomic Dynamics:
  • Production: Follows a Cobb-Douglas function with diminishing returns.
  • Market Imbalance: Defined as $\hat{\phi} = (D-G)/\max(D,G)$, driving price and wage adjustments.
  • Policy: Includes progressive taxation and a modified Taylor rule for interest rates, responding to inflation and unemployment deviations.

3. Key Contributions

1. Unified Simulation System: EconAI is the first LLM-powered system to integrate macroeconomic dynamics and micro-level agent interactions within a single framework, enabling agents to adapt to both market environments and individual economic decisions.
2. Adaptive Mechanisms: The framework introduces the Economic Sentiment Index (ESI), memory weighting, and dynamic decision-making models. These allow the system to continuously adjust to market conditions and simulate real-world fluctuations, surpassing static prediction models.
3. Enhanced Accuracy and Robustness: Through extensive simulations, EconAI demonstrates superior performance in replicating economic indicators (inflation, unemployment) and verifying economic laws (Phillips curve, Okun's law) compared to rule-based, reinforcement learning, and existing LLM-based baselines.

4. Experimental Results

Experiments were conducted over a 20-year horizon using GPT-4o-mini, comparing EconAI against baselines including LEN, CATS, AI-Eco (RL-based), and EconAgent.

• Macro-level Stability: EconAI produced significantly more stable trajectories for inflation, unemployment, and GDP growth. Unlike baselines that exhibited large-amplitude oscillations or degenerate regimes (e.g., AI-Eco reaching 46% unemployment), EconAI's memory and sentiment channels absorbed transient disturbances.
• Economic Regularities: EconAI successfully recovered the textbook downward-sloping Phillips curve (Pearson $\rho = -0.522$) and Okun's law co-movement. Baselines failed to reproduce these relationships, often producing perverse positive slopes due to rigid decision rules.
• Micro-level Dynamics: Simulations of firm interactions successfully captured both destructive competition (price wars, zero-sum outcomes) and healthy competition (innovation, positive-sum growth).
• External Interventions: When injected with a qualitative prompt regarding the 2020 COVID-19 emergency, EconAI agents shifted toward precautionary savings and income diversification, resulting in a realistic spike and lagged recovery in unemployment rates, demonstrating the system's ability to translate qualitative cues into quantitative macro-effects.
• Ablation Studies: Removing historical memory led to sharp unemployment fluctuations; removing the sentiment index increased volatility; removing the belief factor introduced inconsistencies. This confirms the necessity of all components for stability.
• Scalability: Increasing the agent count from 100 to 300 did not significantly alter the stability or realism of inflation rates, indicating robustness to population size changes.

5. Significance and Claims

The paper claims that EconAI represents a crucial step toward realistic, adaptive economic agent simulations. Its primary significance lies in:

• Bridging the Gap: It successfully binds memory, sentiment, and language-level reasoning into a single loop, serving as a viable substitute for hand-engineered rulebooks or bespoke training pipelines.
• Long-term Coherence: By weighting historical memory against current context and modulating decisions via economic sentiment, the system avoids the accumulation of simulation errors common in current LLM-driven approaches.
• Behavioral Realism: The framework captures the dynamic interplay between memory-driven learning and confidence-driven adaptation, essential for accurate long-term economic modeling.

6. Limitations

The authors acknowledge several constraints:

• Computational Overhead: Reliance on large-scale language models incurs significant storage and computational costs, potentially limiting scalability for real-time, extensive simulations.
• Parameter Tuning: Integrating multiple modules (memory decay, sentiment weighting) introduces complexity in parameter tuning, requiring further validation across diverse scenarios.
• Data Scope: Evaluations are primarily based on simulated data, which may not fully capture the nuances of all real-world extreme events or economic fluctuations.