Bringing AI to Life: How Generative Agents Mimic Human Behavior with Memory, Reflection, and Emergent Social Behavior

Two years ago, Stanford and Google researchers released a groundbreaking paper, Generative Agents: Interactive Simulacra of Human Behavior. It explores AI agents designed to simulate lifelike human behaviors. Revisiting this paper, I find their approach—particularly the use of memory and emergent behaviors— worth mentioning.

This blog will focus on the believability of these agents, how they simulate human-like behavior, and how GPT-4 plays a main role in it.

‍

Believable AI Agents

The researchers tested these agents in a sandbox environment called Smallville, a town populated with 25 agents.

The Goal: Illusion of Life

• The agents resemble characters from Disney movies or The Sims.
• They simulate actions like waking up, cooking, working, and socializing.
• Designed to act with apparent volition, creating the facade of realism.

‍

Key Approach: Memory and Reflection

• Memory is central to their architecture.
• Two types of memories:
  1. Observations: Directly perceived events (e.g., "The stove is burning").
  2. Reflections: Higher-level thoughts synthesized from observations (e.g., "I need to be more careful when cooking").
• Reflections allow agents to generalize and infer, much like humans.

‍

How Agents Behave

Memory Stream: The Core

• A database that records all experiences in natural language.
• Helps agents plan actions, react, and maintain long-term coherence.
• Mimics the human brain by filtering irrelevant info and focusing on what’s important.

‍

Memory Retrieval: Intelligent Filtering

To make decisions and respond appropriately, generative agents must retrieve the most relevant memories from their vast memory stream. This process is crucial because not all stored information is equally important or applicable to the current situation. By intelligently filtering memories, the agents can focus on what matters most, ensuring their behavior feels natural and contextually appropriate.

Here’s how the system prioritizes memories:

• Recency: Recent events get priority.
• Importance: Significant events (e.g., breakups) outweigh mundane ones (e.g., breakfast).
• Relevance: Context-specific memories are surfaced (e.g., retrieving study-related memories during a test).

Decision-Making and Reflection

• Reflections guide behavior by synthesizing past experiences.
• Reflections are higher-level, more abstract thoughts generated by the agent. Because they are a type of memory, they are included alongside other observations when retrieval occurs. (interpretation on top of real world observations, or high level observations)
• Example: An agent observes Klaus working on research and Maria doing the same. Reflection helps Klaus realize they share a common interest.
• This mirrors human introspection: the ability to interpret and learn from past events.

‍

A Nod to Westworld: The "Reverie" File

Interestingly, the researchers include a folder named Reverie in their GitHub repository. This name likely references Westworld’s "Reverie" feature, which gave hosts access to latent memories, adding depth to their behavior. Similarly, generative agents use reflections based on past experiences to influence their actions, creating nuanced behavior.

‍

Social Behaviors in Action

With 25 agents living in Smallville, they showed real-life-alike emergent social behaviors:

1. Information Diffusion
   • Agents spread news naturally through conversations.
   • Example: News about Sam’s candidacy for mayor and Isabella’s Valentine’s Day party spread organically.
2. Relationship Memory
   • Agents form and recall social connections.
   • Example: An agent remembers meeting Latoya in the park and later asks about her photography project.
3. Coordination
   • Agents collaborate on group activities.
   • Example: Isabella organizes a Valentine’s Day party, invites others, and decorates the cafe with Maria’s help.

‍

Measurable Results on social behavior

Information Diffusion

As noted in Section 7.1.1, information diffusion is a well-studied phenomenon in social sciences. The researchers expected the agents to spread important news among themselves, and the results confirmed this:

Simulation Results (7.1.2)

• Sam’s Mayoral Candidacy: Awareness increased from 1 agent (4%) to 8 agents (32%).
• Isabella’s Valentine’s Day Party: Awareness increased from 1 agent (4%) to 13 agents (52%).

These results demonstrate the agents’ ability to spread information naturally, just as humans do in real social networks.

‍

Planning and Reacting

‍

Planning for Consistency

• Agents create long-term plans and break them into smaller tasks.
• Example: Klaus plans his day, from studying at the library to taking a walk in the park.

‍

Reacting to Real-Time Events

• Agents adapt their plans based on new observations.
• Example: If a stove starts burning, the agent turns it off and updates their plan.

‍

Evaluation: Testing Believability

The researchers tested the agents’ human-like behavior through interviews. Five key areas were evaluated:

1. Self-Knowledge: Can agents describe themselves?
   • Example: “I’m John, a pharmacist at Willow Market. I love helping my customers.”
2. Memory: Can they recall past events?
   • Example: “Yes, Isabella invited me to the Valentine’s Day party.”
3. Planning: Do they have consistent future plans?
   • Example: “At 10 am tomorrow, I’ll be working at the pharmacy.”
4. Reactions: Can they handle unexpected events?
   • Example: “If my breakfast is burning, I’ll turn off the stove and remake it.”
5. Reflections: Can they synthesize experiences?
   • Example: “Since Wolfgang loves mathematical music, I’ll buy him a book on the subject for his birthday.”

Results showed agents with both observational and reflective memory performed significantly better, exhibiting more human-like behavior.

(I can't help but recall this scene in West World:)

‍

The Role of GPT-4 in the Agents’ Architecture

The generative agents in this study are built on GPT-4, a highly advanced language model which shows exceptional psychological and social reasoning capabilities since 2 years ago. As highlighted in Microsoft’s research paper Sparks of Artificial General Intelligence: Early Experiments with GPT-4, the base model itself possesses remarkable abilities in social intelligence and understanding human behavior. This foundational strength makes the agents appear believable and capable of simulating lifelike interactions.

Psychological Capabilities of GPT-4

Microsoft’s 155-page research paper showcases GPT-4’s advanced psychological reasoning through several experiments, demonstrating its ability to navigate complex social and emotional scenarios. Below are key highlights:

1. False-Belief Understanding (Figure 6.1)

• GPT-4 successfully passes the Sally-Anne false-belief test, a classic experiment in developmental psychology used to assess the ability to understand that others can hold beliefs different from reality.
• This indicates that GPT-4 can reason about others’ mental states and model their perspectives, a fundamental aspect of social intelligence.

2. Emotional Reasoning (Figure 6.2)

• In complex scenarios, GPT-4 demonstrates the ability to infer the emotional states of others.
• Example: GPT-4 reasons about how a character might feel in situations involving interpersonal conflicts or misunderstandings.

3. Intention Recognition (Figure 6.3)

• GPT-4 outperforms earlier models, such as ChatGPT, in understanding the intentions of individuals in nuanced social situations.
• Example: It can accurately interpret why a person might act deceptively or altruistically, showing a deep grasp of human motivations.

Implications for Generative Agents

Given GPT-4’s advanced psychological capabilities, it is no surprise that the generative agents built on this foundation exhibit believable social behaviors. The ability to:

• Model beliefs,
• Reason about emotions, and
• Understand intentions

enables these agents to navigate social dynamics in ways that feel strikingly human. The architecture effectively leverages GPT-4’s inherent social intelligence to simulate interactions such as spreading information, forming relationships, and coordinating group activities.

This strong foundation emphasizes how critical GPT-4’s psychological reasoning is to the overall believability of generative agents in creating the illusion of life.

‍

Emergent Social Behavior among LLM Agents

Another related recent research published in Science Advances titled "Emergent Social Conventions and Collective Bias in LLM Populations" highlights new groundbreaking findings on how large language models (LLMs) autonomously develop social conventions, collective biases, and adapt to social dynamics. These insights are crucial for understanding the behavior of AI agents in multi-agent systems and ensuring their alignment with human values.

Key Insights from the Study

1. Spontaneous Emergence of Conventions:
   • LLM agents can establish universally accepted conventions (e.g., shared linguistic norms) through decentralized, local interactions without explicit programming.
   • This behavior was modeled using the "naming game," where agents reach a global consensus from random initial states.
2. Formation of Collective Bias:
   • Even when individual agents lack inherent biases, collective biases can emerge during convention formation.
   • These biases arise organically from memory and interaction dynamics, leading to certain conventions being favored over others.
3. Adversarial Influence and Critical Mass:
   • Committed minorities of adversarial agents can overturn established conventions if their size reaches a critical mass.
   • The critical threshold depends on factors such as the strength of the existing convention and the architecture of the LLMs.
4. Model-Specific Behaviors:
   • Different LLMs (e.g., Llama-2, Llama-3, Claude-3.5) exhibit varying patterns in forming conventions and resisting adversarial influence.
   • These variations highlight the importance of understanding model-specific dynamics in multi-agent AI systems.

Implications for Generative Agents

This study reinforces the idea that AI agents, particularly those built on advanced LLMs like GPT-4, can exhibit lifelike social behaviors and adapt to complex interactions. The findings emphasize the importance of:

• AI Alignment: Ensuring multi-agent systems align with human values to avoid undesirable norms or harmful biases.
• Social Simulations: Using AI agents to model societal behaviors and address global challenges, such as climate change or public health crises.
• Bias Mitigation: Developing techniques to identify and manage emergent biases in AI populations.

By integrating these insights into generative agents, we can better understand their ability to simulate human-like behaviors, not only at an individual level but also within group dynamics. This positions generative agents as tools for both advancing AI and exploring complex social phenomena.

‍

Final Thoughts

This paper—Generative Agents: Interactive Simulacra of Human Behavior is a milestone in creating AI agents with lifelike behavior. The use of memory—both observational and reflective—is particularly groundbreaking. It mirrors human cognition, where we prioritize, filter, and synthesize experiences to make decisions.

One aspect worth noting is how reflections guide not just the agents’ actions but also their identities. This raises an intriguing question for us: What questions do we ask ourselves? The focus of these questions shapes our stories, much like reflections shape the agents’ behaviors.

For those interested in learning further, I recommend reading the paper and checking out their GitHub repository.

‍

Conclusion: The Value of Being Human

AI has shown remarkable ability to mimic certain aspects of human behavior, yet it cannot replace the core essence of what makes us human. If we view the world solely through the lens of functionality—how different people contribute or produce outcomes—then yes, AI may seem capable of replacing some of these functions. However, humanity is far more than a collection of functions.

As humans, we possess the capacity to love, to care, and to connect—qualities that may not serve a direct functional purpose but lie at the heart of our existence. Love, in particular, is beyond calculation, and it is through love, empathy, and connection that humanity has endured. Our ancestors, even amidst scarcity and conflict, did not merely seek to maximize resources for survival. Instead, they cultivated bonds, built communities, and ensured the survival of the species through cooperation and compassion.

When we look at human history, it becomes clear that humanity has survived countless upheavals—wars, disasters, and moments of massive destruction. Time and again, it is the strength of our shared humanity, our resilience, and our ability to come together during critical moments that has allowed us to persist. This moment in history, with the rapid rise of AI, will require us to adapt once again.

However, we must not fall into the illusion that nothing will change. Structural shifts are inevitable as AI transforms how we live and work. Yet, as long as we retain our humanity—our ability to love, empathize, and unite—we will find a way forward, just as we always have. The future may bring change, but it is up to us to shape that change with the values that define what it means to be human.

‍

‍