What question did this study set out to answer?

The goal is to develop a curiosity-driven AI model that improves learning by addressing memory gaps and conflicts.

April 13, 2026Open Access

Curiosity Beyond Novelty: Conflict, Reward, and Resonance Gaps in Dual-Memory AI

Key Points

The goal is to develop a curiosity-driven AI model that improves learning by addressing memory gaps and conflicts.
Introduced the Resonance Gap Curiosity (RGC) Module within the AURA-X dual-memory architecture.
Formalized curiosity as a three-dimensional construct: Novelty, Conflict, and Reward Curiosity.
Proposed a local memory update rule as an alternative to global backpropagation.
Developed a complete sequential curiosity-learning algorithm and identified six failure modes with mitigations.
Established that Conflict Curiosity aids in belief revision rather than just knowledge accumulation.
Demonstrated that gaps in memory serve as motivational signals for exploration and adaptive learning.
The RGC Module is compatible with existing supervised and self-supervised learning frameworks.

Abstract

Current machine learning systems typically treat novelty as a failure condition: unknown inputs lead to hallucination, low-confidence silence, or out-of-distribution breakdown. Biological learning systems behave differently. Gaps between present perception and stored memory are not merely errors; they function as motivational signals that drive exploration, selective confirmation, and adaptive memory consolidation. This paper introduces the Resonance Gap Curiosity (RGC) Module, a mathematically grounded extension of the AURA-X Ω dual-memory (TM–BM) architecture. The module formalizes curiosity as a three-dimensional construct comprising Novelty Curiosity (Cn), triggered by absent resonance; Conflict Curiosity (Cc), triggered when familiar inputs carry contradictory valence relative to stored memory; and Reward Curiosity (Cr), driven by high-weight positive memory traces. A bounded, contradiction-resistant local memory update rule is proposed as a trace-specific alternative to global backpropagation within the architectural scope of the framework, alongside a complete sequential curiosity-learning algorithm, habituation as the attentional complement to curiosity, and six formally specified failure modes with corresponding mitigations. A central theoretical claim is that Conflict Curiosity constitutes an informationally non-redundant channel beyond base resonance alone, thereby enabling belief revision rather than mere knowledge accumulation. The RGC Module is positioned not as an isolated curiosity mechanism, but as one component within a broader continuity-centered, multi-layer architecture comprising numerous interrelated conceptual and operational layers, including Wajdan, Curiosity, Zameer, memory, safety, moral, and resonance-regulatory processes. The framework is proposed as an architectural overlay compatible with existing supervised and self-supervised systems, and as a formal bridge toward biologically inspired, continual, and self-corrective learning in interactive AI.

Curiosity Beyond Novelty: Conflict, Reward, and Resonance Gaps in Dual-Memory AI

Key Points

Abstract

Cite This Study