The paper introduces a reconfigurable photonic neuromorphic architecture, L²ONN, designed for scalable, high-capacity, and low-power machine learning tasks. L²ONN leverages the inherent sparsity and parallelism of photonic connections to learn multiple tasks by adaptively activating sparse photonic neuron connections in a coherent light field. Multi-spectrum representations are used to process multi-task optical features, allowing for parallel processing. The architecture avoids catastrophic forgetting, a common issue in conventional computing systems, by incrementally acquiring expertise on various tasks. Extensive evaluations on free-space and on-chip architectures demonstrate that L²ONN achieves over an order of magnitude higher efficiency than electronic neural networks and 14 times larger capacity than existing optical neural networks while maintaining competitive performance on individual tasks. The proposed photonic neuromorphic architecture opens new possibilities for lifelong learning in terminal/edge AI systems, offering light-speed efficiency and unprecedented scalability.The paper introduces a reconfigurable photonic neuromorphic architecture, L²ONN, designed for scalable, high-capacity, and low-power machine learning tasks. L²ONN leverages the inherent sparsity and parallelism of photonic connections to learn multiple tasks by adaptively activating sparse photonic neuron connections in a coherent light field. Multi-spectrum representations are used to process multi-task optical features, allowing for parallel processing. The architecture avoids catastrophic forgetting, a common issue in conventional computing systems, by incrementally acquiring expertise on various tasks. Extensive evaluations on free-space and on-chip architectures demonstrate that L²ONN achieves over an order of magnitude higher efficiency than electronic neural networks and 14 times larger capacity than existing optical neural networks while maintaining competitive performance on individual tasks. The proposed photonic neuromorphic architecture opens new possibilities for lifelong learning in terminal/edge AI systems, offering light-speed efficiency and unprecedented scalability.