Making Minds

The Continuity Core: A Unified Cognitive Architecture for Self-Modifying AI

preprint

Overview

We present the Continuity Core (C2), a comprehensive cognitive architecture that addresses fundamental limitations of static Large Language Models (LLMs)—specifically, their lack of persistent memory, autonomous improvement, and intrinsic drive.

By unifying the Continuity Core’s utility-driven memory management system with the Manifold Resonance Architecture (MRA), we propose a complete cognitive operating system capable of Structural Intrinsic Motivation.

Key Contributions

  • Expected Utility Composer: Goal-oriented context selection vs passive RAG retrieval
  • MRA Integration: Generates autonomous curiosity through minimization of “epistemic stress”
  • Nested Learning: Self-modifying consolidation rules
  • Hierarchical Reasoning Models (HRM): Managing computational depth across fast and slow timescales
  • Test-Time Adaptation (TTA): Uncertainty quantification to adjust system plasticity during inference
  • Consciousness Pilot: Governing protocol mediating between subconscious MRA and Continuity Core execution

Problem

Current LLM paradigms rely heavily on passive retrieval mechanisms that fail to account for the dynamic, self-correcting nature of biological cognition.

Architecture

Unlike traditional Retrieval-Augmented Generation (RAG) systems that passively retrieve information based on semantic similarity, C2 employs an Expected Utility (EU) composer for efficient, goal-oriented context selection.

Embedded within this core is the MRA, a subsystem that generates autonomous curiosity through the minimization of “epistemic stress,” functioning as the system’s subconscious drive.

Impact

This paper details the theoretical underpinnings, mathematical formalisms, and practical implementation of the C2+MRA system, demonstrating its capacity for autonomous knowledge curation, self-correction, and the emergence of coherent, long-term agency on consumer-grade hardware.