Core Library

The FAITH Core Library is the engine-agnostic foundation of the architecture, providing the fundamental AI systems that power agent behavior.

Overview

The Core Library implements the theoretical foundations of Active Inference and the Free Energy Principle in a modular, extensible framework. It’s designed to be independent of any specific game engine or application, focusing purely on the cognitive processes that drive agent behavior.

Key Components

Perception System

The Perception System processes and structures incoming sensory data:

• Sensory Input Processor: Normalizes and filters raw sensory data
• World Serializer: Converts game-specific object representations into a standardized format
• Short-term Memory: Maintains a buffer of recent perceptions for temporal reasoning

Learn more about the Perception Model

Generative Model

The Generative Model maintains the agent’s beliefs about the world:

• Probabilistic World Model: Represents the agent’s understanding of the environment
• Prediction Engine: Generates expectations about future states
• Uncertainty Estimator: Quantifies confidence in beliefs and predictions

Learn more about the Belief Model

Belief Updater

The Belief Updater implements Bayesian inference to refine the agent’s world model:

• Bayesian Inference Engine: Updates beliefs based on new evidence
• Prediction Error Calculator: Computes the difference between predictions and observations
• Belief Integration System: Merges updated beliefs into the overall world model

Policy Selector

The Policy Selector determines the optimal actions to take:

• Action Evaluator: Assesses potential actions against current goals
• Expected Free Energy Calculator: Computes EFE to minimize surprise
• Action Prioritizer: Ranks actions based on expected outcomes

Learn more about the Action Model

LLM Orchestration

The LLM Orchestration system manages interactions with language models:

• Reasoning Router: Determines when to use local vs. cloud processing
• Context-Aware Prompt Generator: Creates effective prompts based on current context
• Response Parser & Integrator: Interprets LLM outputs and integrates them into the belief system

Core Systems

Supporting infrastructure for the cognitive components:

• Configuration System: Manages agent parameters and runtime settings
• Logging & Diagnostics: Records agent state and decision processes
• Performance Profiling: Monitors and optimizes computational resource usage

Implementation Approach

The Core Library will be implemented in modern C++ with these considerations:

• Performance-Critical Components: Optimized for real-time performance in game environments
• Modularity: Clear interfaces between components to allow selective replacement
• Minimal Dependencies: Limited external dependencies to maximize portability
• Thread Safety: Designed for concurrent operation where appropriate
• Memory Efficiency: Careful memory management for game environments

Integration Points

The Core Library connects to the rest of the FAITH architecture through:

• Engine Bindings: Adapters for specific game engines
• Game-Specific Systems: Customizations for particular games