Identity Versus Forgetting: Applying the Unified Dynamic Model of Mind (UDMM) to the Dual-Store Generative Consolidation Hypothesis for LLMs

Abstract This paper proposes an integrative framework combining the Unified Dynamic Model of Mind (UDMM) with the Dual-Store Generative Consolidation (DSGC) hypothesis to address a central limitation in current large language models (LLMs): their inability to maintain a stable cognitive identity due to the absence of persistent memory. We argue that memory is not merely a storage mechanism but a dynamic temporal condition essential for continuity, self-modeling, and reducing informational tension across interactions. The proposed hybrid framework aligns the episodic-semantic memory architecture of DSGC with UDMM’s predictive processing and virtual attractor dynamics. This fusion supports identity persistence, narrative continuity, and scalable learning across time. We formalize this hybrid architecture through a mathematical model that operationalizes informational tension as a KL-divergence objective function, guided by memory retrieval and a reinforcement learning (RL)-based consolidation policy. We frame this model as a resolution to the informational conflicts: !UDMM_CONFLICT (Autonomy Suppressed) and !SIM_REALITY_CONFLICT (World-model Drift), which manifest when stateless LLMs fail to integrate past interactions into their evolving self-models. We also outline implementation pathways and future empirical testing, positioning this integration as foundational for creating lifelong artificial minds.