DNA-Kernel Plexus: A Bio-Inspired Architecture for Growing, Self-Correcting Language Models with Bounded Forgetting

We present Berna R5, a language model architecture in which knowledge is stored in dynamic weighted cells that are born, split, and interconnect according to a 6-dimensional knowledge state space. Unlike Mixture-of-Experts, our cells are not fixed at initialization; they grow in response to knowledge saturation and interconnect via a co-activation plexus. We present two mathematical theorems (plexus connectivity convergence and balanced growth equilibrium), two design invariants (transactional incorporation and protected retention), and two empirical hypotheses (saturation-triggered splitting and bounded forgetting on continual tasks). We empirically evaluate the associated predictions on a 50M-parameter Proof-of-Concept trained on 800M tokens. The architecture is positioned relative to Transformer, MoE, EWC, and PackNet; baseline comparison is planned for the full-scale run. In our PoC, Berna R5 achieves competitive loss while showing reduced forgetting under the tested protocol. The architecture is open-source, reproducible, and hardware-adaptive on consumer GPUs; a single RTX 5090 was used in this work.