RDLNN-DCRAS: Non-Destructive Continual Learning via Hebbian Lateral Voting in a Biologically-Grounded Liquid Neural Cognitive Architecture

We present DCRAS (Distributed Coupled Reservoir Attractor System), the cross-columnar
voting and continual learning subsystem of the Recurrent Dendritic Liquid Neural Network
(RDLNN) — a 10-module biologically-grounded cognitive architecture integrating Liquid
State Machines, Amygdala fear conditioning, Basal Ganglia action selection, and Hippocampal episodic memory. DCRAS eliminates catastrophic forgetting for sequential associative
learning by avoiding the mechanism that causes it: unlike gradient-based approaches that
overwrite shared weight configurations, DCRAS stores associative mappings as attractors
in fixed-weight LSM reservoirs coupled via Hebbian lateral voting matrices. New knowledge
accumulates additively in high-dimensional orthogonal subspaces rather than competing
destructively for shared parameters.
We validate this claim at three reservoir scales using a fully vectorized batched CUDA
engine (B=8 seeds in parallel, Tesla T4). In an 8-task sequential aversive conditioning
paradigm (CS-A through CS-H, 25 trials each), RDLNN-DCRAS achieves mean retention
0.984 (±0.044) at N=512, and perfect 1.000 (±0.000) at both N=1024 and N=2048,
with Backward Transfer BWT=+0.000 at the two larger scales. All 8 seeds pass the ≥0.875
threshold at every scale. A blank-input saturation probe (CS-null) scores 0.000 avoidance
across all 24 probe measurements (8 seeds × 3 scales), ruling out non-specific excitation.
An LSTM baseline exhibits BWT=−0.458; LSTM+EWC achieves BWT=−0.048.
We further demonstrate that RDLNN-DCRAS maintains latent representations of absent
sensory modalities through continuous cross-modal constraint propagation, achieving 0.817
cosine similarity to the reference attractor under complete sensory deprivation — without
generative training, diffusion models, or autoregressive loss. A control agent with voting
disabled collapses to 0.287. These results establish RDLNN-DCRAS as a non-destructive
continual learning system operating in a fundamentally different regime from gradient-based
sequential learners, with scale-invariant zero-forgetting confirmed across a 16× range of
reservoir size.