Paper deep dive
EconAI: Dynamic Persona Evolution and Memory-Aware Agents in Evolving Economic Environments
Annie Liu, Zane Cao, Lang Chen, Zongxin Xu, Zigan Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/8/2026, 3:16:53 PM
Summary
EconAI is an LLM-powered simulation framework that integrates economic sentiment indexing, memory weighting, and dynamic decision-making to model macroeconomic and microeconomic environments. It enables adaptive, human-like agent behaviors by balancing short-term market signals with long-term strategic planning, outperforming traditional rule-based and machine learning baselines in economic stability and realism.
Entities (17)
Relation Signals (17)
EconAI → simulates → Households
confidence 95% · Our framework categorizes agents into two types: households for microeconomic analysis and firms for macroeconomic perspectives.
EconAI → simulates → Firms
confidence 95% · Our framework categorizes agents into two types: households for microeconomic analysis and firms for macroeconomic perspectives.
EconAI → uses → Large Language Models
confidence 95% · The integration of large language models (LLMs) in economic simulations has significantly enhanced agent-based modeling
EconAI → incorporates → Economic Sentiment Index
confidence 92% · incorporating economic sentiment indexing (ESI), memory weighting, and dynamic decision-making mechanisms.
Economic Sentiment Index → influences → Agent Decision-Making
confidence 92% · Incorporating the dynamic changes in economic belief makes the decision-making of the intelligent agent not only data-dependent but also emotionally influenced.
Memory Weighting → adjusts → Historical Data Influence
confidence 90% · By quantifying economic belief, adjusting historical data influence, and linking work-consumption behaviors, EconAI achieves a more human-like decision process
EconAI → incorporates → Dynamic Decision-Making
confidence 90% · incorporating economic sentiment indexing (ESI), memory weighting, and dynamic decision-making mechanisms.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The integration of large language models (LLMs) in economic simulations has significantly enhanced agent-based modeling, yet existing frameworks struggle to capture the interplay between short-term optimization and long-term strategic planning. Conventional approaches rely on static data-driven predictions, failing to incorporate adaptive behaviors influenced by economic sentiment, market volatility, and individual goals. To address these limitations, we introduce a novel EconAI framework, incorporating economic sentiment indexing (ESI), memory weighting, and dynamic decision-making mechanisms. By quantifying economic belief, adjusting historical data influence, and linking work-consumption behaviors, EconAI achieves a more human-like decision process, where agents adapt their actions based on both market signals and long-term objectives. It is the first LLM-powered simulation system that can simulate the macro/microeconomic environment and interactions in a unified framework. Empirical evaluations show that EconAI improves stability in economic responses, better replicates real-world employment-consumption cycles, and enhances overall decision robustness. This advancement marks a crucial step towards more realistic, adaptive economic agent simulations.
Tags
Links
- Source: https://arxiv.org/abs/2605.13762v1
- Canonical: https://arxiv.org/abs/2605.13762v1
Trouble viewing inline? Open PDF directly →
Full Text
42,590 characters extracted from source content.
Expand or collapse full text
EconAI: Dynamic Persona Evolution and Memory-Aware Agents in Evolving Economic Environments Annie Liu, Zane Cao, Lang Chen, Zongxin Xu, Zigan Wang Tsinghua University Abstract The integration of large language models (LLMs) in economic simulations has significantly enhanced agent-based modeling, yet existing frameworks struggle to capture the interplay between short-term optimization and long-term strategic planning. Conventional approaches rely on static data-driven predictions, failing to incorporate adaptive behaviors influenced by economic sentiment, market volatility, and individual goals. To address these limitations, we introduce a novel EconAI framework, incorporating economic sentiment indexing (ESI), memory weighting, and dynamic decision-making mechanisms. By quantifying economic belief, adjusting historical data influence, and linking work-consumption behaviors, EconAI achieves a more human-like decision process, where agents adapt their actions based on both market signals and long-term objectives. It is the first LLM-powered simulation system that can simulate the macro/microeconomic environment and interactions in a unified framework. Empirical evaluations show that EconAI improves stability in economic responses, better replicates real-world employment-consumption cycles, and enhances overall decision robustness. This advancement marks a crucial step towards more realistic, adaptive economic agent simulations. EconAI: Dynamic Persona Evolution and Memory-Aware Agents in Evolving Economic Environments Annie Liu, Zane Cao, Lang Chen, Zongxin Xu, Zigan Wang Tsinghua University 1 Introduction The advent of artificial intelligence (AI) Huang et al. (2026); Yin et al. (2026); Cao et al. (2024) has revolutionized economic research, enabling unprecedented insights into individual behaviors, consumer preferences, and market dynamics through advanced data processing and pattern recognition Jorgenson (2001); Schorfheide and Song (2015); Christiano et al. (2005). Among AI-driven methodologies, Agent-based modeling (ABM) has emerged as a powerful framework for simulating economic systems from the bottom up, capturing interactions among diverse agents without relying on predetermined equilibria Farmer and Foley (2009). While early ABM approaches relied on simplistic rule-based models Tesfatsion and Judd (2006); Brock and Hommes (1998), recent advancements have incorporated learning-based techniques to better reflect complex economic dynamics Trott et al. (2021); Zheng et al. (2022); Mi et al. (2023). However, tailoring decision-making processes to individual agents remains a significant challenge. Customized rule sets require extensive expert knowledge and calibration Windrum et al. (2007), while neural network-based approaches often face prohibitive computational costs and training complexities Mi et al. (2023). These limitations hinder the practical application of ABM and restrict its ability to capture the full diversity of economic dynamics. The integration of large language models (LLMs) into economic simulations has significantly advanced agent-based modeling by enhancing reasoning and planning capabilities Zhao et al. (2023b). However, existing frameworks often fall short in capturing the nuanced interplay between short-term optimization and long-term strategic planning, which is critical for realistic economic decision-making. A key limitation lies in their inability to effectively integrate two essential components: event memory and economic beliefs. In real-world economic environments, decision-makers rely on historical experiences (event memory) and adapt their actions based on evolving market conditions and their confidence in the economic outlook. Current LLM-driven approaches, however, often neglect these aspects, leading to an accumulation of simulation errors that can render the model increasingly unrealistic over time Yue et al. (2024). This gap highlights a critical limitation of current methods: their inability to simulate the dynamic interplay between memory-driven learning and confidence-driven adaptation, which is essential for accurate long-term economic modeling. In light of these limitations, this paper addresses the following question: How can we develop a new system that can more accurately simulate macro/microeconomic dynamics in a unified framework? To this end, we introduce EconAI, a novel framework designed to enhance the long-term decision-making capabilities of LLMs in economic simulations. Our framework categorizes agents into two types: households for microeconomic analysis and firms for macroeconomic perspectives. It also integrates the roles of government and financial institutions, recognizing their substantial impact on macroeconomic conditions. EconAI incorporates a knowledge memory module for each agent type, constructed from historical interactions and knowledge learned through LLMs. This memory is organized into long-term and short-term banks, capturing enduring economic patterns and immediate market conditions, respectively. The long-term memory bank stores vector representations of high-level event summaries, while the short-term memory bank retains contextual information for ongoing activities. A dedicated event memory perception module helps maintain coherence across sessions by effectively managing these memory banks, ensuring decisions are grounded in both historical context and current conditions. Furthermore, EconAI introduces a persona extraction module that models the evolving preferences and economic beliefs of agents. These dynamic personas are continuously updated and stored in a long-term persona bank, allowing for personalized and context-sensitive decision-making. By integrating long-term memory, dynamic persona modeling, and real-time updates, EconAI enables agents to respond to immediate economic fluctuations while considering past experiences and adjusting their strategies according to evolving economic beliefs. Our experiments demonstrate that EconAI significantly improves the accuracy of traditional economic indicators such as inflation and unemployment rates, outperforming conventional rule-based or machine-learning agents. This highlights the framework’s ability to balance short-term responsiveness with long-term strategic thinking, providing a more robust platform for economic simulations. In summary, our contributions are three-fold: • We introduce EconAI, the first LLM-powered simulation system that integrates both macroeconomic dynamics and micro-level agent interactions within a unified framework. This allows agents to adapt to both market environments and individual economic decisions effectively. • EconAI incorporates multiple adaptive mechanisms, including the Economic Sentiment Index (ESI), memory weighting, and dynamic decision-making models. These features enable it to continuously adjust to market conditions and simulate real-world economic fluctuations, surpassing traditional static prediction models. • Through extensive macroeconomic and microeconomic simulations, we demonstrate that EconAI significantly outperforms existing models, and EconAI simulates economic indicators such as inflation and unemployment with greater precision and verifies different economic laws. 2 Related Work 2.1 Simulation in Macroeconomics Macroeconomic modelling has long been caught between analytical tractability and behavioural realism. Closed-form frameworks such as DSGE Christiano et al. (2005) and reduced-form empirical models Hendry and Richard (1982); Phelps (1967); Kydland and Prescott (1982) impose strong rationality and equilibrium assumptions, which limits their capacity to represent heterogeneous agents, frictions, and out-of-equilibrium dynamics. Agent-Based Modelling (ABM) sidesteps this constraint by letting macro-level regularities emerge from the interaction of many micro-level actors, and has therefore become the preferred vehicle for policy experiments whose outcomes depend on distributional shifts rather than representative-agent averages. Within the ABM tradition, two design philosophies dominate. Hand-crafted rule-based agents Tesfatsion and Judd (2006); Brock and Hommes (1998) encode economists’ prior knowledge directly but scale poorly as the behavioural space grows, requiring expert re-calibration for every new setting. Learning-based agents trained with deep or reinforcement learning Trott et al. (2021); Zheng et al. (2022); Mi et al. (2023) relax this rigidity but transfer the burden to data and compute, neither of which is easy to come by at the macro scale. EconAI departs from this dichotomy by grounding agent cognition in language-level reasoning, enabling a single backbone to absorb macroeconomic context and act heterogeneously across households and firms without hand-written rules or task-specific training. 2.2 LLM-empowered Agents A rapidly growing literature turns large language models into autonomous simulation agents Wang et al. (2023); Xi et al. (2023), exploiting their ability to follow natural-language instructions, retain conversational context, and recover gracefully from under-specified inputs Team (2022); Yoheinakajima (2023). These capabilities have been put to use in generative-agent societies Park et al. (2023a, 2022), multi-agent social-science testbeds Kovač et al. (2023); Gao et al. (2023); Jinxin et al. (2023); Gilbert and Troitzsch (2005), and tool-using scientific assistants Boiko et al. (2023); Bran et al. (2023). The economic applications that are closest to our setting range from probing the rationality of individual LLM decisions Horton (2023); Chen et al. (2023b), through game-theoretic interaction between small numbers of agents Guo (2023); Akata et al. (2023), to market-scale simulation with competing firms or traders Zhao et al. (2023a); Chen et al. (2023a). The work most directly comparable to ours is EconAgent Li et al. (2024), which uses an LLM to drive single-period household choices within a macroeconomic loop. Our focus is complementary: rather than treating each month as an independent rational choice, EconAI models the longitudinal trajectory of a household, tightly coupling memory retention, belief revision, and sentiment-modulated preferences inside one decision module, so that multi-year shocks and regime shifts are absorbed rather than reset at every step. Figure 1: The illustration of the simulation for the microeconomic and macroeconomic environments (left) and our EconAI (right). On the left, the simulation analyzes household microeconomic decisions on work and leisure, alongside firms’ macroeconomic choices on production, investment, and employment, shaped by simplified interactions with government and financial institutions. On the right, the event module retains historical data in long-term memory and contextual information in short-term memory. The persona module dynamically extracts and updates user and agent personas from ongoing interactions, storing them in a dedicated personal bank. The response module integrates this information to generate contextually informed and coherent responses, enabling the system to support economic decision-making and simulate human-like behaviors such as self-management, cooperation, and competition. 3 EconAI Framework In this section, we introduce EconAI, which simulates the economic activity plan in the decision-making process. This plan comprises a series of abstract actions to be carried out across various scenarios with a long-term-based framework. According to economists Falk et al. (2018), long-term decision-making rather than short-term counterparts is the base for a robust economic system. To this end, as illustrated in Figure 1, we develop a new approach to simulate the long-term decision-making process including event perception, dynamic personas extraction, and response generation for decision-making, thus conducting the economic simulation. 3.1 Event Perception The overall pipeline is illustrated in Figure 1. We now turn to the agent-side architecture that drives each household and firm inside EconAI. For an agent, such as a household, it has specified metadata such as the profession, specialty, skills, credentials, and experiences of the agent. The agent observes information from the environment, makes decisions, and conducts the appropriate action. In real-world economic activities, humans often make decisions based on heuristic rules and established customs derived from experience, such as the belief that lower bank interest rates encourage investment. Afterward, people reinforce and apply these rules of thumb based on past successes or revise their heuristic rules in response to failures in specific contexts. Much like humans, the agent’s brain serves as a central nucleus driven by an LLM. The brain module enables the agent to exhibit sophisticated cognitive abilities critical for professional-grade performance, including memory, planning, and reasoning. Memory Storage. The event memory module is designed to perceive historical events to generate coherent responses across interval time. As shown in Figure 1, this event memory module is segmented into two major sub-modules that focus separately on long-term and short-term memory. The memory module is designed to capture and encode events from previous interactions in economic activities. This process involves logging both the timestamps t and concise summaries o, which are transformed into representations and stored in a low-cost memory bank ML=ϕ(tj,oj)∣j∈1,2,…,lM_L=\φ(t_j,o_j) j∈\1,2,…,l\\. Here, ϕ(⋅)φ(·) denotes the text encoder (such as MiniLM Wang et al. (2020)), while l represents the size of the memory bank. The encoded information is then made easily accessible via an embedding-based retrieval mechanism, facilitating efficient access to stored memories. Event Summary. Unlike previous agent methods Park et al. (2023b) that depend solely on the zero-shot capabilities of LLMs for event extraction and summarization, we incorporate instruction tuning Wei et al. (2022) into our event summary module to enhance the quality of the summaries for economic activities. We remember the event in a structured format consisting of: (1) a task background introduction, (2) relevant economic activities to be comprehended, and (3) specific summarization instructions. These elements serve as input prompts and generate the output. The event summary module is fine-tuned using this data, effectively improving its summarization capabilities. 3.2 Economic Preference and Confidence Memory information is objective, but human decision-making is often affected by emotions and confidence, such as being overly pessimistic during an economic recession and overly optimistic during an overheated economy. Equal weighting of short-term and long-term information cannot reflect people’s different sensitivities to economic changes in different time frames. Then, we introduce the Economic Sentiment Index (ESI) to quantify the economic belief of agents and store this ESI value in the memory module: ESIt=λESIt−1+(1−λ)ESIt,LLMESI_t= _t-1+(1-λ)ESI_t,LLM (1) Where λ is the memory decay factor (0.8 0.95). ESIt,LLMESI_t,LLM is the latest economic sentiment index generated by LLM. In this way, the economic sentiment of each quarter can be quantified and stored for subsequent decision-making calculations. In this way, we can make the results of LLM reflection quantifiable, reducing computational overhead. Incorporating the dynamic changes in economic belief makes the decision-making of the intelligent agent not only data-dependent but also emotionally influenced. The impact of historical economic data will decay, avoiding excessive reactions by the intelligent agent due to short-term economic shocks. Then, we use confidence to adjust the final decision. If confidence is low (e.g., below 0.5). Adjust the final decision in conjunction with the economic sentiment index: pwi=σ(pwi−θESIt),pci=σ(pci+βESIt)p_w_i=σ(p_w_i- _t),p_c_i=σ(p_c_i+ _t) (2) If economic belief is low (ESIt<0ESI_t<0), the intelligent agent reduces consumption and increases the willingness to work. If economic belief is high (ESIt>0ESI_t>0), the intelligent agent increases consumption and reduces work. 3.3 Economic Environment Design This section outlines the design and decision-making processes of intelligent economic agents, government-imposed taxation policies, production-consumption dynamics, and financial market mechanisms shaping macroeconomic behavior. These components interact within a unified framework that captures real-world economic principles. Agent Decision-Making Framework. The core decision-making process for each agent involves two key economic choices: work and consumption. Initially, these decisions are modeled as: pwt,pct∼LLM(zi,P,si,u,r,ESIt)p_w_t,p_c_t (z_i,P,s_i,u,r,ESI_t) (3) where pwtp_w_t is the labor market participation probability, and pctp_c_t is the consumption-to-income ratio. pwip_w_i denotes the agent’s willingness to work, determining whether they choose employment, while pcip_c_i represents the agent’s consumption propensity, indicating the proportion of income allocated to spending. The variable ziz_i corresponds to the agent’s individual income in the current period, while P reflects the prevailing market price of essential goods. sis_i accounts for the agent’s current savings, which influence financial security and spending capacity. The unemployment rate u provides insight into labor market conditions, affecting job availability and economic confidence. Lastly, r represents the bank interest rate, which impacts savings growth and borrowing costs. These inputs collectively allow the LLM to generate dynamic and context-aware decisions that align with real-world economic behaviors. Production and Consumption Dynamics. Labor supply contributes to goods production following a Cobb-Douglas function: G←G+S=G+AKtβ1Ltβ2e−λKt,G← G+S=G+AK_t _1L_t _2e^-λ K_t, (4) where A represents the overall productivity factor, KtK_t is the capital stock, and Lt=∑j=1Nlj×168L_t= _j=1^Nl_j× 168 denotes the total labor hours supplied. The diminishing returns parameter λ ensures that excessive capital accumulation does not indefinitely increase output. Capital evolves dynamically as: Kt=(1−δ)Kt−1+I−ωKt,K_t=(1-δ)K_t-1+I-ω K_t, (5) where δ represents depreciation, I is new investment, and ωKtω K_t accounts for inefficiencies. Aggregate goods demand is determined byD=∑j=1Ndj=∑j=1NqjcsjPD= _j=1^Nd_j= _j=1^N q_j^cs_jP, where qjcq_j^c represents the consumption propensity of agent j, sjs_j is the savings level, and P is the market price of goods.Then, the market imbalances is defined as: φ^=D−Gmax(D,G) = D-G (D,G). It drives price and wage adjustments: wi←wi(1+φi⋅κ),P←P(1+φP⋅κ),w_i← w_i(1+ _i·κ), P← P(1+ _P·κ), (6) where φi∼sign(φ^)U(0,βw|φ^|) _i ( )U(0, _w| |) influences wages, and so does φp _p. Government Taxation and Redistribution. To enhance economic stability and equity, the government implements a progressive taxation system, where tax rates rise with income. Post-tax income incorporates redistributed tax revenues, promoting partial wealth reallocation and reducing extreme income inequality. In financial markets, agents’ savings grow at an interest rate influenced by unemployment and inflation dynamics. This rate follows a modified Taylor rule, adjusting for deviations from target inflation and natural unemployment. Detailed formulations and mechanisms are provided in the Appendix. 3.4 Response Generation When the agent engages in economic activity, it will recall the memory module, which combines event extractor: retrieved relevant memories m, the short-term memory context MSM_S, and the personas PaP_a representing the agent. These inputs are passed to a response generator, which produces an appropriate response r. To improve the agent’s capability to generate coherent and contextually relevant responses for economic activities, for each example spanning a series of times (such as several months in the simulation), we dynamically simulate the economic activity’s progression. As new economic events arise, the system applies event summarization, persona extraction, and topic-aware memory retrieval to collect relevant context and retrieve associated memories and personas for the user and agent. This rich data is then incorporated into a response generation prompt (details in Appendix). On top of this, the agent can summarize the satisfied score for the current state and then make the decisions. 4 Experiments In this section, we conduct experiments to study the ability of EconAI, aiming to answer the following research questions (RQ). RQ1: How does EconAI perform in simulations compared to conventional models? RQ2: What role do the key components of EconAI play in influencing the simulation outcomes? RQ3: Is the decision-making process within EconAI interpretable, and do simulations effectively capture the impact of external interventions? 4.1 Experimental Setup Baselines. We benchmark EconAI against four representatives of the current design space. LEN Lengnick (2013) and CATS Gatti et al. (2011) are long-established rule-based macro-ABMs whose hand-crafted work/consumption rules have been repeatedly validated in follow-up simulation studies, making them natural yard-sticks for the rule-based end of the spectrum. To approximate the heterogeneity that real populations exhibit, we additionally report a Composite variant in which each synthetic agent draws its rule from the union of the LEN and CATS rulebooks at initialisation and holds it fixed thereafter. On the learning-based side, AI-Eco, based on AI-Economist Zheng et al. (2022), trains utility-maximising policies via reinforcement learning Arulkumaran et al. (2017); it represents the optimisation-driven end of the spectrum. Finally, we compare against EconAgent Li et al. (2024), the closest LLM-based baseline, which drives per-period household decisions through a perception/reflection loop; including it allows us to isolate the contribution of EconAI’s memory-weighting and sentiment-indexing mechanisms. Definition of Economic Indicators. The annual nominal GDP is calculated as the total of S×PS× P over the course of a year. To compute real GDP, we set the first year of the simulation as the base year, replacing P with P0P_0, where P0P_0 represents the price of goods in that reference year. The wage inflation rate follows a similar formula to price inflation, except the average price is substituted by the average wage across all agents. For households, disposable income refers to the total income left after taxes and essential expenditures. The savings rate is the proportion of disposable income that is saved rather than used for consumption. For firms, profit margin is defined as the ratio of net profit to total revenue, reflecting the firm’s profitability. Simulation Setup. Every synthetic household is initialised with a short persona string (a name sampled from the LLM, an age, and an occupation) that is injected verbatim into the prompt the agent receives at each step, giving the model a stable reference point across months. Ages are drawn from the 2018 U.S. Census adult distribution restricted to the 18 to 60 working population Bureau (2024), while hourly earnings follow a Pareto distribution whose scale and shape parameters are calibrated so that the induced monthly income distribution and effective tax brackets match 2018 U.S. tax filings Zheng et al. (2022). To inject wage dispersion into the textual layer of each persona, we ask the LLM to propose ten representative job titles per income decile and sample an occupation accordingly. Employment transitions are sticky: a household that worked in month t keeps its job at t+1t+1, whereas an unemployed household is presented with a fresh offer whose wage is resampled from the current cross-sectional distribution. Full distributional statistics are deferred to the supplementary material. All experiments use GPT-4o-mini through the OpenAI API, with simulation orchestration implemented in Python. 4.2 Macro-level Analysis (RQ1) (a) (b) (c) (d) Figure 2: Annual variations of macroeconomic indicators, where the simulation based on EconAI shows more stable and numerically plausible indicators. (a) (b) Figure 3: Economic regularity study. Economic Indicators. Figure 2 tracks four aggregate indicators, namely annual inflation, unemployment, nominal GDP, and nominal-GDP growth, over the 20-year horizon. AI-Eco drifts into a degenerate regime, with its unemployment rate pinned near 46% and nominal output swinging over implausibly wide bands, so we drop it from the plots rather than let it compress the remaining curves. The rule-based and reinforcement-learning baselines share a common failure mode: every indicator shows repeated large-amplitude oscillations, a symptom of decision rules that cannot smooth short-run shocks before they propagate through the labour and consumption loop. EconAI’s trajectories sit in a substantially tighter envelope without any per-variable tuning, because the memory and sentiment channels absorb transient disturbances inside the agent rather than routing them back into aggregate demand. We read this as the qualitative signature of a simulator whose micro-level decisions approximate how human households actually adjust under uncertainty. (a) (b) Figure 4: The experiments about the macro-level analysis. Economic Regularity. A stronger test of macroeconomic realism is whether simulated aggregates reproduce canonical empirical regularities. We evaluate two: the Phillips relation between unemployment and wage inflation Phelps (1967), and Okun’s law linking the change in unemployment to real-output growth Okun (1963). Figure 3 reports both. EconAI recovers the textbook downward-sloping Phillips relation (Pearson ρ=−0.522ρ=-0.522, p<0.01p<0.01) and the expected co-movement between unemployment growth and real-GDP growth under Okun’s law. EconAgent and the rule-based and learning-based baselines fail at least one of the two tests, and the rule-based family in particular produces a perverse positive Phillips slope, an artefact of decision rules that do not condition consumption on the employment signal. We attribute EconAI’s success to the behavioural channel inside the agent: households contract spending when their perceived unemployment risk rises, and this precautionary response propagates through aggregate demand to wages rather than being hard-coded as a slope. 4.3 Micro-level Analysis (RQ1) (a) (b) (c) Figure 5: More experimental studies. In the economic environment modeled by EconAI, simulations of firms reveal classic market strategies and economic phenomena such as competition and cooperation. Our experiments are illustrated in Figure 4, One scenario where one firm’s revenue stagnates or declines significantly suggests destructive competition. This occurs when firms engage in aggressive price wars, excessive resource allocation to outcompete rivals or strategic behaviors that undermine industry stability. In game theory, such dynamics resemble a zero-sum or negative-sum game, where one firm’s gain comes directly at the expense of another, sometimes leading to a market exit or inefficiencies. When competition becomes excessively aggressive without value creation, the industry as a whole may suffer from reduced profitability and instability. This simulation thus provides an economically rational representation of competition, reinforcing that constructive rivalry leads to mutual growth, while destructive competition can hinder overall economic progress. In contrast, in the second scenario, where both firms experience revenue growth, the simulation reflects principles of healthy competition. Economic theory suggests that when firms compete in ways that drive innovation, improve efficiency, and expand market demand, both can thrive. This aligns with the game-theoretic concept of positive-sum competition, where firms engage in strategic differentiation rather than direct conflict, allowing them to capture different market segments or create complementary value. By fostering innovation and efficiency rather than resorting to destructive tactics, both firms achieve sustainable long-term growth, leading to increased overall welfare in the industry. 4.4 Ablation Study (RQ2) We conduct an ablation study by independently removing key economic factors, with the results over a 10-year span shown in Figure LABEL:Ablation_study. Specifically, we analyze five versions: the original model (EconAI), removing historical memory (denoted as No History Memory), removing the sentiment index (No Sentiment Index), removing the belief factor (No Belief Factor), and removing investment impact (No Investment Impact). Notably, the absence of historical memory leads to sharp fluctuations in the unemployment rate, suggesting that past economic trends play a crucial role in stabilizing employment levels. Similarly, removing the sentiment index results in increased volatility, highlighting the influence of public sentiment on labor market dynamics. The absence of the belief factor introduces inconsistencies, reinforcing its role in guiding employment expectations. Finally, removing investment impact leads to overall lower unemployment variability, suggesting that investment plays a stabilizing role. These findings emphasize the importance of incorporating these components. 4.5 External Intervention (RQ3) To probe whether EconAI responds plausibly to an exogenous regime shift, we inject a single textual event (the March 2020 declaration of a COVID-19 national emergency in the United States) into every agent’s prompt from that month onward and leave every other parameter untouched Dawid and Gatti (2018). Figure LABEL:COVID-19 contrasts the resulting unemployment trajectory against a counterfactual run without the injection. Simulated unemployment spikes sharply in early 2020, recovers only partially, and remains elevated relative to the counterfactual through 2021, consistent with the lagged labour-market recovery documented by Organization for Economic Co-operation and Development (1970), even though the agents are given nothing more quantitative than a qualitative cue. Inspection of the reflection traces shows households shifting toward precautionary savings and reasoning explicitly about income diversification, which is the behavioural channel through which the prompt is converted into aggregate effects. We therefore read the experiment as evidence that the prompt interface is a sufficient conduit for macro-level shocks; numerical exactness is not the intended deliverable. Sensitivity and Robustness We increase the number of agents to 300 and rerun the simulation. As illustrated in Figure LABEL:Inflation_rate_under_different_number_of_agents, the inflation rates remain consistently stable and realistic, closely resembling the results obtained with 100 agents. This trend is also observed across other economic indicators, indicating that the simulation outcomes are not significantly affected by changes in the number of agents. 5 Conclusion We presented EconAI, an LLM-driven agent-based macroeconomic simulator whose household and firm sides share a single cognitive backbone composed of event perception, a weighted long-term/short-term memory bank, and a sentiment-modulated response module. The key departure from prior LLM-agent work is the Economic Sentiment Index together with an explicit weighting between remembered history and current context, which together let an agent modulate its work–consumption decisions according to its evolving belief about the environment rather than a static rule or a single-period optimum. Over a 20-year simulation, this design is what keeps inflation, unemployment, and nominal GDP inside empirically plausible envelopes; it is also what lets EconAI recover the downward Phillips slope and Okun’s co-movement on which the rule-based and learning-based baselines fail, and convert a single qualitative COVID-19 prompt into a quantitatively coherent labour-market response. The broader take-away is that binding memory, sentiment, and language-level reasoning into one loop is a viable substitute for hand-engineered rulebooks or bespoke training pipelines in macroeconomic agent design. 6 Limitation The proposed EconAI framework demonstrates promising advancements by integrating event memory, economic sentiment, and dynamic persona extraction to bridge short-term optimization and long-term strategic planning. However, several limitations remain. First, its reliance on large-scale language models results in significant computational and storage overhead, potentially impeding scalability and real-time applicability in extensive simulations. Second, the complexity of integrating multiple modules introduces challenges in parameter tuning, with key parameters such as the memory decay factor and sentiment weighting requiring further validation across diverse economic scenarios. Finally, the experimental evaluations are primarily based on simulated data, which may not fully capture the nuances of all real-world economic fluctuations and extreme events. Addressing these limitations will enhance the model’s robustness and broaden its practical applicability. References E. Akata, L. Schulz, J. Coda-Forno, S. J. Oh, M. Bethge, and E. Schulz (2023) Playing repeated games with large language models. arXiv preprint arXiv:2305.16867. Cited by: §2.2. K. Arulkumaran, M. P. Deisenroth, M. Brundage, and A. A. Bharath (2017) Deep reinforcement learning: a brief survey. IEEE Signal Processing Magazine 34 (6), p. 26–38. Cited by: §4.1. D. A. Boiko, R. MacKnight, and G. Gomes (2023) Emergent autonomous scientific research capabilities of large language models. arXiv preprint arXiv:2304.05332. Cited by: §2.2. A. M. Bran, S. Cox, A. D. White, and P. Schwaller (2023) ChemCrow: augmenting large-language models with chemistry tools. arXiv preprint arXiv:2304.05376. Cited by: §2.2. W. A. Brock and C. H. Hommes (1998) Heterogeneous beliefs and routes to chaos in a simple asset pricing model. Journal of Economic dynamics and Control 22 (8-9), p. 1235–1274. Cited by: §1, §2.1. U. C. Bureau (2024) Access demographic, economic and population data from the u.s. census bureau.. External Links: Link Cited by: §4.1. Z. Cao, J. Li, Z. Wang, and J. Li (2024) Diffusione: reasoning on knowledge graphs via diffusion-based graph neural networks. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 222–230. Cited by: §1. J. Chen, S. Yuan, R. Ye, B. P. Majumder, and K. Richardson (2023a) Put your money where your mouth is: evaluating strategic planning and execution of llm agents in an auction arena. arXiv preprint arXiv:2310.05746. Cited by: §2.2. Y. Chen, T. X. Liu, Y. Shan, and S. Zhong (2023b) The emergence of economic rationality of gpt. arXiv preprint arXiv:2305.12763. Cited by: §2.2. L. J. Christiano, M. Eichenbaum, and C. L. Evans (2005) Nominal rigidities and the dynamic effects of a shock to monetary policy. Journal of political Economy 113 (1), p. 1–45. Cited by: §1, §2.1. H. Dawid and D. D. Gatti (2018) Agent-based macroeconomics. Handbook of computational economics 4, p. 63–156. Cited by: §4.5. A. Falk, A. Becker, T. Dohmen, B. Enke, D. Huffman, and U. Sunde (2018) Global evidence on economic preferences. The quarterly journal of economics 133 (4), p. 1645–1692. Cited by: §3. J. D. Farmer and D. Foley (2009) The economy needs agent-based modelling. Nature 460 (7256), p. 685–686. Cited by: §1. C. Gao, X. Lan, Z. Lu, J. Mao, J. Piao, H. Wang, D. Jin, and Y. Li (2023) S3: social-network simulation system with large language model-empowered agents. arXiv preprint arXiv:2307.14984. Cited by: §2.2. D. D. Gatti, S. Desiderio, E. Gaffeo, P. Cirillo, and M. Gallegati (2011) Macroeconomics from the bottom-up. Vol. 1, Springer Science & Business Media. Cited by: §4.1. N. Gilbert and K. Troitzsch (2005) Simulation for the social scientist. McGraw-Hill Education (UK). Cited by: §2.2. F. Guo (2023) GPT agents in game theory experiments. arXiv preprint arXiv:2305.05516. Cited by: §2.2. D. F. Hendry and J. Richard (1982) On the formulation of empirical models in dynamic econometrics. Journal of Econometrics 20 (1), p. 3–33. Cited by: §2.1. J. J. Horton (2023) Large language models as simulated economic agents: what can we learn from homo silicus?. Technical report National Bureau of Economic Research. Cited by: §2.2. Y. Huang, J. Chen, S. Zhang, and Z. Cao (2026) RaDAR: relation-aware diffusion-asymmetric graph contrastive learning for recommendation. In Proceedings of the ACM Web Conference 2026, p. 6445–6456. Cited by: §1. S. Jinxin, Z. Jiabao, W. Yilei, W. Xingjiao, L. Jiawen, and H. Liang (2023) Cgmi: configurable general multi-agent interaction framework. arXiv preprint arXiv:2308.12503. Cited by: §2.2. D. W. Jorgenson (2001) Information technology and the us economy. American Economic Review 91 (1), p. 1–32. Cited by: §1. G. Kovač, R. Portelas, P. F. Dominey, and P. Oudeyer (2023) The socialai school: insights from developmental psychology towards artificial socio-cultural agents. arXiv preprint arXiv:2307.07871. Cited by: §2.2. F. E. Kydland and E. C. Prescott (1982) Time to build and aggregate fluctuations. Econometrica: Journal of the Econometric Society, p. 1345–1370. Cited by: §2.1. M. Lengnick (2013) Agent-based macroeconomics: a baseline model. Journal of Economic Behavior & Organization 86, p. 102–120. Cited by: §4.1. N. Li, C. Gao, M. Li, Y. Li, and Q. Liao (2024) Econagent: large language model-empowered agents for simulating macroeconomic activities. In ACL, p. 15523–15536. Cited by: §2.2, §4.1. Q. Mi, S. Xia, Y. Song, H. Zhang, S. Zhu, and J. Wang (2023) TaxAI: a dynamic economic simulator and benchmark for multi-agent reinforcement learning. arXiv preprint arXiv:2309.16307. Cited by: §1, §2.1. A. M. Okun (1963) Potential gnp: its measurement and significance. Cowles Foundation for Research in Economics at Yale University. Cited by: §4.2. Organization for Economic Co-operation and Development (1970) Infra-Annual Labor Statistics: Unemployment Rate Total: From 15 to 64 Years for United States. External Links: Link Cited by: §4.5. J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein (2023a) Generative agents: interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442. Cited by: §2.2. J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein (2023b) Generative agents: interactive simulacra of human behavior. In ACM UIST, p. 2:1–2:22. Cited by: §3.1. J. S. Park, L. Popowski, C. Cai, M. R. Morris, P. Liang, and M. S. Bernstein (2022) Social simulacra: creating populated prototypes for social computing systems. In Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology, p. 1–18. Cited by: §2.2. E. S. Phelps (1967) Phillips curves, expectations of inflation and optimal unemployment over time. Economica, p. 254–281. Cited by: §2.1, §4.2. F. Schorfheide and D. Song (2015) Real-time forecasting with a mixed-frequency var. Journal of Business & Economic Statistics 33 (3), p. 366–380. Cited by: §1. A. Team (2022) AutoGPT: the heart of the open-source agent ecosystem. Note: (Accessed on 01/10/2023)https://github.com/Significant-Gravitas/AutoGPT Cited by: §2.2. L. Tesfatsion and K. L. Judd (2006) Handbook of computational economics: agent-based computational economics. Elsevier. Cited by: §1, §2.1. A. Trott, S. Srinivasa, D. van der Wal, S. Haneuse, and S. Zheng (2021) Building a foundation for data-driven, interpretable, and robust policy design using the ai economist. arXiv preprint arXiv:2108.02904. Cited by: §1, §2.1. L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y. Lin, et al. (2023) A survey on large language model based autonomous agents. arXiv preprint arXiv:2308.11432. Cited by: §2.2. W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, and M. Zhou (2020) MiniLM: deep self-attention distillation for task-agnostic compression of pre-trained transformers. In NeurIPS, Cited by: §3.1. J. Wei, M. Bosma, V. Y. Zhao, K. Guu, A. W. Yu, B. Lester, N. Du, A. M. Dai, and Q. V. Le (2022) Finetuned language models are zero-shot learners. In ICLR, Cited by: §3.1. P. Windrum, G. Fagiolo, and A. Moneta (2007) Empirical validation of agent-based models: alternatives and prospects. Journal of Artificial Societies and Social Simulation 10 (2), p. 8. Cited by: §1. Z. Xi, W. Chen, X. Guo, W. He, Y. Ding, B. Hong, M. Zhang, J. Wang, S. Jin, E. Zhou, et al. (2023) The rise and potential of large language model based agents: a survey. arXiv preprint arXiv:2309.07864. Cited by: §2.2. M. Yin, Z. Cao, Q. Xia, C. Tu, and N. Gao (2026) NumCoKE: ordinal-aware numerical reasoning over knowledge graphs with mixture-of-experts and contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 16172–16180. Cited by: §1. Yoheinakajima (2023) BabyAGI. Note: (Accessed on 01/10/2023)https://github.com/yoheinakajima/babyagi Cited by: §2.2. Z. Yue, L. Zhang, and Q. Jin (2024) Less is more: mitigating multimodal hallucination from an EOS decision perspective. In ACL, p. 11766–11781. Cited by: §1. Q. Zhao, J. Wang, Y. Zhang, Y. Jin, K. Zhu, H. Chen, and X. Xie (2023a) Competeai: understanding the competition behaviors in large language model-based agents. arXiv preprint arXiv:2310.17512. Cited by: §2.2. W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang, Z. Dong, et al. (2023b) A survey of large language models. arXiv preprint arXiv:2303.18223. Cited by: §1. S. Zheng, A. Trott, S. Srinivasa, D. C. Parkes, and R. Socher (2022) The ai economist: taxation policy design via two-level deep multiagent reinforcement learning. Science advances 8 (18), p. eabk2607. Cited by: §1, §2.1, §4.1, §4.1.