Liquid Latent Synthesis (LLS-4): A Quad-Partite Architecture for Latent-Space Concept Interpolation in Continual Reinforcement Learning

Catastrophic forgetting remains a central obstacle to deploying continually-learning agentson resource-constrained edge hardware. Existing mitigations regularization-based methodssuch as Elastic Weight Consolidation (EWC) and architecture-growth methods such as Progressive Neural Networks either accumulate numerically unstable penalty terms as the number oftasks grows, or scale parameter count linearly with the number of skills learned. We proposeLiquid Latent Synthesis (LLS-4), an architecture that decouples fast online policy learning from oine concept consolidation. A Closed-form Continuous-time (CfC) liquid networkencodes physical experience into a compact 16-dimensional latent manifold during an awakephase; during an oine sleep phase, a synthesis engine identies pairs of latent memoriesthat are temporally proximate but semantically distinct, linearly interpolates them, decodes theinterpolated vector into synthetic training data, and distills the result into a lightweight (tensof-thousands-of-parameter) adapter module routed at inference time by a soft-attention contextgate. We implement the full pipeline in PyTorch and ncps, and report a rst pilot comparisonagainst a standard LSTM and a custom diagonal-Fisher EWC baseline on a synthetic two-taskregression proxy. The pilot run shows that the current prototype's shared backbone is notprotected during sequential training, producing a retention score (4.5%) below both baselines(33.5% and 49.2%) the opposite of the architecture's design intent. A secondary zero-shotsynthesis metric (71.0%) is also reported, but we identify and disclose a degeneracy in its construction that likely inates this number independent of any genuine skill combination. Wetreat both ndings as diagnostic rather than conrmatory, and use them to motivate a concrete,itemized experimental protocol backbone freezing, parameter-matched baselines, multi-seedevaluation, a sleep-phase ablation, and a non-degenerate zero-shot task required before anyclaim of solving catastrophic forgetting can be supported. This paper's contribution is thereforethe architecture, an open, working prototype, and a transparent account of what remains to bevalidated.