Pearl is a production-ready reinforcement learning agent library from Meta’s Applied Reinforcement Learning team. Where most RL code is research-grade and fragile, Pearl is engineered for real systems: agents that optimize long-term cumulative reward and keep working under partial observability, sparse feedback, and high stochasticity.
Pearl is open source on GitHub, where it has earned more than 3,000 stars under the MIT license, with development continuing into 2026. It was highlighted on Meta’s official NeurIPS 2023 channels and is accompanied by a paper and an official project site with detailed introductions. The library provides a modular architecture covering the pieces production RL systems need, policy learners, exploration strategies, replay mechanisms, and history summarization for partially observable environments, designed to be composed rather than rewritten for every new problem. It addresses settings from contextual bandits up to full reinforcement learning, which matches how applied teams actually work: many production problems are bandit-shaped, and Pearl lets you start there and grow. Because it comes from a team that runs RL at Meta scale, the emphasis is on correctness, performance, and APIs that survive contact with real data pipelines, not just benchmark leaderboards. For researchers it offers clean reference implementations; for practitioners it offers a path from prototype agents to systems that can be deployed, monitored, and improved over time.
Pearl is free and open source under the MIT license. There is no hosted service or paid tier; you run it within your own training and serving infrastructure.
Pearl is for reinforcement learning researchers and applied ML engineers building adaptive systems, recommendation, ranking, and decision-making products among them, who need RL code that is meant to leave the notebook and run in production.