DMN v3.0 - The Dream Loop: Memory Consolidation, Familiarity-Weighted Retrieval, and the Learned Graph Executor in the Lár Default Mode Network

DMN v3.0 adds three mechanisms to the Lár Default Mode Network architecture: Consolidation Loop, Familiarity-Weighted Retrieval, and Learned Graph Executor. The Consolidation Loop mirrors biological memory consolidation with three tiers—Episodic, Semantic, and Procedural—and requires exclusive, isolated cycles to avoid interference. Familiarity-Weighted Retrieval uses a saturating, temporally-decaying confidence measure for memory retrieval, grounded in dual-process theory. The Learned Graph Executor distinguishes the routing policy from Mixture-of-Experts by using offline imitation learning, analogous to hierarchical reinforcement learning's Options Framework. A prospective hybrid with a JEPA world model enables one-step look-ahead planning. DMN v2.0 is deployed; v3.0 is a formal specification.

Study at a glance

Abstract

We present DMN v3.0, a formally specified extension to the Lár Default Mode Network (DMN) architecture introducing three novel architectural mechanisms: the Consolidation Loop, Familiarity-Weighted Retrieval, and the Learned Graph Executor (LGE). The Consolidation Loop defines a three-tier memory progression — Episodic (ChromaDB), Semantic (Dreamer-generated narratives), and Procedural (LoRA weight-delta training) — that mirrors the hippocampal-to-cortical consolidation pathway observed in biological memory systems. We establish a biological constraint derived from in-vitro neural network research (van der Kolk & van Pelt, 2024) that consolidation must run in exclusive, isolated cycles, preventing interference with active task processing. Familiarity-Weighted Retrieval replaces unbounded cosine similarity scoring with a bounded, saturating recognition signal R(r, t_c) = 1 − exp(−β · r · exp(−λt_c)), grounded in dual-process memory theory (Mandler, 1980; Yonelinas, 2002). This provides a principled, temporally-decaying confidence measure for memory retrieval that prevents stale memories from dominating context. The Learned Graph Executor formally distinguishes the DMN BrainNode routing policy from Mixture-of-Experts (MoE) architectures by framing it as an offline imitation-learned meta-controller, structurally analogous to the Options Framework in Hierarchical Reinforcement Learning. We further specify a prospective Lár-JEPA BrainNode hybrid where a JEPA world model performs one-step look-ahead planning over candidate routing actions before an LLM executes the selected LGSL instruction. All mechanisms are specified as prospective prior art. The DMN v2.0 implementation is stable and deployed; DMN v3.0 is a formal architectural specification authored and published prior to any employment or corporate collaboration. This preprint extends a prior art chain of five published Zenodo records (DOIs: 10.5281/zenodo.19025925, 19120047, 19245328, 19484646, 19516414).