This article introduces a photonic neuromorphic architecture, L²ONN, designed for lifelong learning across tens of tasks. The architecture leverages the inherent sparsity and parallelism of photonic connections to enable efficient, scalable, and low-power learning. Unlike traditional electronic systems, which face energy and scaling limitations, L²ONN avoids catastrophic forgetting by adaptively activating sparse photonic neuron connections and gradually increasing activation for new tasks. It processes multi-task features in parallel using multi-spectrum representations allocated with different wavelengths. The system demonstrates superior performance compared to existing optical and electronic neural networks, achieving over an order of magnitude higher efficiency and 14× larger capacity while maintaining competitive task-specific accuracy. The architecture is validated through both free-space and on-chip implementations, showing its ability to learn and adapt to new tasks with high accuracy. The on-chip version, fabricated using standard CMOS technology, offers a compact, scalable solution for terminal/edge AI systems. The proposed photonic neuromorphic architecture provides a new approach to lifelong learning, enabling efficient, scalable, and versatile AI systems. The study highlights the potential of photonic computing to overcome the limitations of electronic systems, offering a promising solution for real-world applications requiring high performance and energy efficiency.This article introduces a photonic neuromorphic architecture, L²ONN, designed for lifelong learning across tens of tasks. The architecture leverages the inherent sparsity and parallelism of photonic connections to enable efficient, scalable, and low-power learning. Unlike traditional electronic systems, which face energy and scaling limitations, L²ONN avoids catastrophic forgetting by adaptively activating sparse photonic neuron connections and gradually increasing activation for new tasks. It processes multi-task features in parallel using multi-spectrum representations allocated with different wavelengths. The system demonstrates superior performance compared to existing optical and electronic neural networks, achieving over an order of magnitude higher efficiency and 14× larger capacity while maintaining competitive task-specific accuracy. The architecture is validated through both free-space and on-chip implementations, showing its ability to learn and adapt to new tasks with high accuracy. The on-chip version, fabricated using standard CMOS technology, offers a compact, scalable solution for terminal/edge AI systems. The proposed photonic neuromorphic architecture provides a new approach to lifelong learning, enabling efficient, scalable, and versatile AI systems. The study highlights the potential of photonic computing to overcome the limitations of electronic systems, offering a promising solution for real-world applications requiring high performance and energy efficiency.