4.2.1 Core Architecture Design

Core Architecture Design

The agent framework consists of multiple modules that work in close coordination to support the perception, decision-making, and action capabilities of agents. According to the architecture diagram, the agent framework includes the following core components:

1. World Rules Module

• This includes trading systems, housing systems, education systems, and other modules related to the operation of a virtual society. These rules define the basic operational logic of the environment in which agents exist and constrain their behavior.

• For example, the trading system enables economic activities for agents, and the election system supports democratic decision-making among agents.

2. Behavior and Planning Module

• The Planner is the core module for agent behavior, encompassing the following functionalities:

• Environmental Perception: Captures environmental data through sensors or virtual inputs, identifying task objectives and external changes.

• Daily Planner: Generates daily activity plans for agents, dynamically adjusting behavior based on time and resources.

• Detailed Planner and Behavior Sequence Generator: Further refines task planning, generating specific behavior sequences for the execution module.

• The behavior planning module integrates with the Event Queue to enable dynamic task distribution and priority management.

3. Memory and State Management Module

• Running State serves as the memory module for agents, storing the current task state and long-term memory data. It supports behavior optimization through the following functionalities:

• Short-Term Memory: Enables real-time task scheduling and state feedback.

• Long-Term Memory: Helps agents retain experiences and recall relevant information for future decisions.

• The Memory Module also facilitates data sharing among internal submodules (e.g., Action Agent and Conversation Agent), improving overall efficiency.

4. Dialogue Module and Natural Language Processing

• The dialogue module enables agents to interact with users or other agents through natural language, including the following core features:

• Message Distribution: Distributes user inputs or environmental information to the relevant modules for processing.

• Time and Topic Planner: Determines the logic and content direction of conversations based on context.

• Agents can dynamically adjust task behaviors through dialogue, enhancing interaction with users.

5. Advanced Abstraction Modules

• Agents possess advanced modules such as emotion, psychological states, and constitutions, making their behavior more human-like and complex.

• These modules support the long-term behavioral evolution of agents and the establishment of relationships with other agents through continuously updated state management.