In addition to the intentional introduction of information into the state with documents and integrations, interacting with the model adds signals to the state definition through your prompts, responses, reactions to responses, and clarifications.

Unlike the more easily reviewed inputs, this information is largely inferred from the overall conversation. Explicit commands like “give a one word answer” will affect the immediate output, but repeated requests and patterns will establish themselves within the state over interactions.

I refer to this conversational information as the Me Domain. It includes user preferences, habits, communication style, conversation goals, and inferred characteristics. It is accumulated from multiple sources, but functions somewhat like a cheat sheet for the interaction.

For instance, if you use a window frequently and solely for generating recipes, the system may infer without explicitly being told that you like to cook. The inference goes beyond just the frequency of the recipes, but includes the way you request and respond in the conversation. If you regularly show excitement and enthusiasm over suggested recipes, the system may infer enjoyment. If you introduce each request with “ugh, I have to cook again,” it will likely infer the opposite.

Interactions are building and refining the current state.

While these state adjustments can be subtle, they shape the responses based on those preferences. In the aforementioned example, if the AI assumes you like to cook, recipes may be more complex. If it believes you don’t like cooking, the recipes may tend to become simpler with shorter prep times.

Interactions are building and refining the current state. You can therefore steer the state by conscious and purposeful input designed to refine the environment. Many prompt templates that are available are doing just this. They are defining the rules of engagement for the state. These can be effective at the beginning of a conversation to establish a starting state, used throughout the conversation to continue guiding and refining the state, or even applied to response specific requests.

At a practical level, the Me Domain is important to state control as it is the model’s record of conversation state assumptions. Beyond this useful function, the Me Domain may also explain why some characteristics that appear human can emerge from models.