SOLAR proposes a parameter-level meta-learning framework that treats model weights as an explorable environment for autonomous adaptation in non-stationary streams. The arXiv preprint (v1, 23 Mar 2026) details a multi-level RL loop that consolidates commonsense priors then discovers modification strategies stored in an evolving episodic buffer to balance plasticity and stability. Experiments report gains over baselines across commonsense, math, medical, coding, social, and logical tasks. Related work on elastic weight consolidation (Kirkpatrick et al., PNAS 2017) and progressive neural networks (Rusu et al., 2016) highlighted similar stability-plasticity tensions but relied on fixed architectures rather than open-ended self-optimization. SOLAR extends these by removing manual curation and gradient-heavy fine-tuning, addressing gaps in short-horizon agent benchmarks that the primary source only implicitly contrasts.