A Modular Cognitive Architecture for Autonomous, Self-Motivated Agents

We present a practical, modular architecture for autonomous, self-motivated cognitive agents capable of continuous perception, planning, action, and online learning without relying on user prompts to initiate cognition. The architecture integrates a Global Workspace (GW) for attention management, a hierarchical Recurrent Generative World Model (RGWM) for multi-scale internal simulation, intrinsic motivation drives with learning-progress-gated curiosity, hierarchical planning with compute budgets, and provenance-aware memory systems. Unlike traditional prompt-driven language models, this system maintains internal state over time, generates tasks via intrinsic drives, and learns continuously from both experience and human interaction. The design prioritizes autonomy, generalization across domains, continual learning with safety constraints, transparency through causal tracing, and safety-by-construction through policy gates and audit mechanisms. We describe the complete system architecture, present implementation-ready specifications for each component, and outline evaluation methods for measuring functional autonomy, cognitive signatures, and safety properties.