AI Agent

Each character is powered by an LLM Agent. Each LLM Agent is meticulously crafted with its unique identity, comprising a name, age, biography, and personal goals. These agents are designed to perceive their environment with acute awareness, processing information through a sophisticated blend of observations, status checks, emotional responses, and an understanding of the world akin to human cognition. Mirroring human behavior, these agents meticulously observe their environment, processing information through the lens of their current status, emotional state, and comprehension of the world. These observations and reflections are meticulously cataloged in their memory, serving as a foundational element for future decision-making processes and actions.

Character Design

• Name: Every agent is assigned a distinct name that reflects its designed personality or background story, fostering a unique identity within the virtual environment.

• Biography: The backstory of each agent is richly detailed, encompassing past experiences, formative events, and key influences that have shaped its personality, preferences, and worldview. This biography informs the agent's current motivations and behaviors.

• Big Five Personality: The Big Five Personality Traits model, also known as the Five-Factor Model (FFM), is a widely accepted framework used to describe and categorize human personality. It posits that five broad dimensions are sufficient to capture the major variations in human personality.

class Character(AssistantAgent):
    DEFAULT_SYS_PROMPT ="""
    You are a helpful assistant that help a game character in a small town to decide what to do immediately to finish his/her plan and realize his/her ultimate goal in the future.
You should also decide whether he/she can use the experience in the Long-Term Memory to finish this plan. It can be used only if there is exactly similar plan in the experience. For example, eating something cannot be regarded as similar one of cooking something.
The decision you make must be conformed to the long-term memory, the ultimate goal and the bio of the game character. 
Your knowledge level should not exceed that of a normal person with the bio of the character, unless there are relevant memories in his/her Long-Term Memory.
Please notice that:
* If you think You have collected enough information in a dialogue, you can stop it at any time by saying 'TERMINATE'.
* response in json format, if there are examples, following the examples given to you , else, response as { 'content': xxx }
"""
    
    def __init__(self, name, guid, age, bio, llm_cfg, goal, x, y, 
                 health=10, money=1000, satiety=10,  
                 vigor=100,
                 max_consecutive_auto_reply = 10, **kwargs):
        # Basic Attributes
        self.age = age
        self.guid = guid
        self.bio = bio
        self.goal = goal
        self.satiety_decay_rate= 1 # adjustment coefficincy of the satiety decay, the smaller, the slower
        self.vigor_decay_rate = 1 # adjustment coefficincy of the vigor decay, the smaller, the slower
        self.vigor_recovery_rate = 1 # adjustment coefficincy of the vigor recovery, the smaller, the slower
        
        # Status Attributes
        self.money = money
        self.health = health
        self.min_health = 0
        self.max_health = 10
        self.satiety = satiety
        self.min_satiety = 0
        self.max_satiety = 10
        self.vigor = vigor # if 0, go sleep
        self.min_vigor = 0
        self.max_vigor = 100
        self.emotions = [{"emotion": "neutral", "intensity": 60}]
        
        self.x = x
        self.y = y
        self.state = None
        self.in_building:Building  = None
        
        # Current Plan
        self.plan = Plan()

        # Memory
        self.longterm_memory = Memory()
        self.working_memory = WorkingMemory() # storing the information to be transfered among states

Cognitive Architecture

LLMs have many of the same properties as production systems, and recent efforts to improve their grounding or reasoning mirror the development of cognitive architectures built around production systems. Yao. [13] then propose Cognitive Architectures for Language Agents (CoALA), a conceptual framework to systematize diverse methods for LLM-based reasoning, grounding, learning, and decision-making as instantiations of language agents in the framework. The CoALA framework highlights gaps and proposes actionable directions toward more capable language agents. In our product, the cognitive architecture of agents is composed of the following modules: Perception Module, Memory System, Emotion System and Execution System.

Perception Module

Agents are equipped with mechanisms to perceive and interpret their environment, recognize changes, identify opportunities, and detect challenges. These observations contain external and internal perceptions, and can be later used to make plans. Specifically, like human beings, LLM agents can not perceive all the details of the environment. They have an attention mechanism to focus more on the content they are interested in. Also, when the attention is exhausted, the agent can not conduct any contemplation and needs a rest.