I introduce Topology-Rewiring Neural Cognitive Diffusion (TR-NCD), a sequence modeling framework that redefines what a “weight” is and relocates long-term knowledge away from dense floating-point parameter matrices. In TR-NCD, I treat long-term knowledge as a structural graph (G=(V,E,m)) whose edges encode integer-valued masses/counts rather than real-valued multiplicative weights. I treat “weights” at runtime as addressed, ephemeral activation artifacts—identifiers that can be materialized into a temporary working set and then evicted immediately after use. This results in a clean separation between (i) structural parameters (graph topology + integer masses) as durable knowledge and (ii) context state (e.g., an SSM-like module) as temporal control logic for retrieval and ordering, rather than a repository of learned dense numeric weights. I formalize TR-NCD through a set of numbered definitions that specify: the symbol alphabet (including relation atoms), stable identifiers, ephemeral weight materialization, graph-based long-term storage, a thresholded diffusion operator, and training as local topology rewiring rather than backpropagated gradient updates on floating-point matrices. In later sections of the paper (continued in subsequent parts), I extend the model with Expandable Relation Keys (ERK) for variable-order relational contexts and an Ambiguity Decision Manager (ADM) that explicitly handles uncertainty via multi-hypothesis, deferred commitment, and an “ask/expose” mode instead of forced single-path output. I also provide complete algorithmic specifications for inference, training, on-demand creation, and maintenance/pruning, as well as a concrete evaluation plan and ablation matrix template.
Ahmed Hadi Hussein (Thu,) studied this question.