The Macaron-V1 framework introduces a co-designed system for open continual learning, utilizing a Mixture-of-LoRA (MoL) architecture to enable experiential intelligence in real-world environments. Instead of modifying the base model, the MoL approach freezes the underlying parameters and composes specialized Low-Rank Adaptation (LoRA) adapters, selecting a specific adapter for each user turn. The flagship Macaron-V1-Venti model scales to 748B parameters, utilizing a 744B GLM-5.2 base integrated with four specialized LoRAs for chat, agentic tasks, coding, and GenUI. For local deployment, the Macaron-V1-Tall variant offers a 50B parameter configuration based on Qwen3.6-35B. The system achieves adaptation through a recursive self-improvement loop involving versioned model-harness pairs, where experience is evaluated against an external contract via the Harness Context Protocol. Key technical components include the UI4A component-native GenUI harness, the MindForge agentic reinforcement learning (RL) framework, and the LongStraw long-context RL method. The infrastructure is supported by the MinT post-training platform and stability techniques designed for sparse Mixture-of-Experts (MoE) and DSA base models. While current evaluations across Personal Intelligence and GenUI benchmarks validate the system, the researchers note that the compounding gains from collective intelligence and continuous learning remain subjects for ongoing investigation.