Continual Learning via Generational Parameter Transmission: Weight Retention Dynamics in Neural Agents Under Non-Stationary Environments

Catastrophic forgetting remains a primary roadblock in continual learning when deep neural architectures encounter sequential task distributions. In biological and cultural transmission, generational succession offers an evolutionary alternative: lineage adaptation occurs through lossy parameter inheritance rather than indefinite weight preservation. In this work, we extend Generational Artificial Life (GAL) from stylized scalar agents to neural network agents parameterized by Multi-Layer Perceptrons (MLP). Agents undergo gradient-based learning on non-stationary, alternating classification tasks (Circle vs. XOR decision boundaries) under metabolic parameter budgets (||W||_2 <= c) and generational weight transfer (W_child = f * W_parent). Key Findings: 1. High-Volatility Collapse: When task distributions shift rapidly (every 5 generations), neural populations avoid catastrophic interference by actively suppressing retention (f = 0.312 ± 0.330) to reset obsolete feature extractors. 2. Intermediate Stability Basin: At intermediate timescales (10-25 generations), retention stabilizes near f = 0.506 - 0.532, preserving general feature extractors while providing sufficient plastic capacity for fast fine-tuning (test accuracy: 73.5%). Complete source code, multi-seed sweeps, and reproduction scripts: https://github.com/psycho-prince/Matrix Interactive Simulator: https://psycho-prince.github.io/Matrix/