Researchers have introduced RSIAgent, a training-free multi-agent framework designed to facilitate recursive self-improvement in novel environments through autonomous memory construction. The framework utilizes a coordinated architecture consisting of curriculum, actor, and verifier agents to explore unknown interfaces and tools. This process allows the system to capture and retain environment-specific knowledge, specifically focusing on reusable causal relationships between actions, conditions, and consequences. To optimize exploration, RSIAgent employs a "broad-then-deep" strategy: parallel broad recursive self-exploration is used to map diverse environmental structures, while focused deep self-exploration targets complex edge cases, hidden constraints, and unknown causal dependencies. Notably, the resulting memory is frozen and can be utilized for downstream tasks without requiring further model parameter updates. Empirical evaluations conducted on the OSWorld-v2 and Agent's Last Exam benchmarks demonstrate the framework's efficacy. The implementation enables open-source models, such as Kimi-K3 and GLM-5.3, to achieve performance levels that surpass frontier closed-source models, including GPT-6, highlighting the potential of autonomous memory-driven adaptation for digital agents.