Generative Agents
Generative Agents are computer characters controlled by a language program that remember what they experience, draw conclusions from it, and make daily plans. They became known in 2023 through a village simulation in which 25 such characters lived together without a fixed script.
Generative Agents are artificial characters in a computer world that generate their own behavior. They are controlled by a program that has learned to continue text – the same principle that chatbots like ChatGPT are based on. Unlike characters in a normal computer game, they don’t follow a pre-written script. Instead, they receive a short description of their persona, such as occupation, place of residence, and preferences, and then decide step by step what to do on their own. The term became known in 2023 through a research project at Stanford University. There, 25 such characters lived in a simulated village called Smallville, went to work, met at the café, and even organized a Valentine’s Day party that no one had programmed in advance.
What Smallville reveals about believable behavior
What was interesting about the experiment was not the graphics, which looked like an old video game. What was interesting was that social behavior emerged on its own. One character decided to throw a party. She invited others, they spread the word, and in the end several guests showed up at the right place at the right time. This chain of information, transmission, and action was in no program code.
For the gaming industry, this is a big promise. To this day, minor characters in games mostly repeat the same lines over and over. Characters that remember previous conversations and adjust their behavior would make a game world feel significantly more alive. Several companies are working on exactly that, but so far mainly in demos and small test projects.
The second reason for the interest lies in research. Social scientists can use such simulations to play out how rumors spread or how groups react to new rules. This does not replace a real study with humans. But it does provide hypotheses that can then be tested. What remains important here is the researchers' own warning: the characters imitate human behavior, they are not proof of how humans actually act.
Memory, reflection, and daily planning
The core of a Generative Agent is a memory stream. Everything the character experiences is stored as a short sentence, together with the timestamp. One such entry reads, for example: 'Isabella wipes the table at the café.' After a few days, there are thousands of entries. Because a language model cannot process everything at once, a selection must be made of what currently matters.
This selection is made based on an assessment using three criteria. How recent is the memory, how important was the event, and how well does it fit the current situation? The best matches end up in the text that is sent to the language model. You can imagine it like an assistant who, before every conversation, pulls out the matching notes from a huge folder.
In addition, there are two further building blocks. During reflection, the character regularly condenses many individual memories into general conclusions, such as: 'Klaus is very interested in research.' During planning, this results in a rough daily schedule, which is then broken down into individual actions. If the character meets someone along the way, the plan is adjusted. It is precisely this loop of remembering, reflecting, and planning that distinguishes a Generative Agent from a simple chatbot, which starts from scratch again after every response.
From research village to real products
Generative Agents cannot yet be bought directly. The term appears mainly in trade articles, tech news, and announcements from game companies. Nvidia has been showing demos for several years featuring game characters that respond freely. On social networks, there have been experiments with fictitious profiles that post and comment independently.
Related but not the same are AI agents in everyday work. These book appointments, search websites, or write code, and are measured by their results. Generative Agents, on the other hand, are meant to seem believable, not to be productive. Both share building blocks such as memory and planning, but pursue different goals.
The limitations are clear. Every thinking step costs a call to the language model, and thus money. A simulation with 25 characters over two days incurred costs in the four-figure dollar range at Stanford. In addition, the characters are easily influenced and polite to the point of being unbelievable. And when such agents appear on social networks, the question arises whether it must be possible to recognize them as machines.