EvoLib: Turning experience into evolving knowledge

EvoLib is a framework from Microsoft Research that allows AI systems to learn from experience during inference without ground-truth labels or model updates. Unlike traditional memory systems that store raw experiences statically, EvoLib transforms past attempts into reusable skills and reflective insights. These knowledge units are continuously consolidated with similar existing knowledge to become more general, and each unit is weighted based on its long-term impact on future tasks. The framework was tested on mathematical reasoning, code efficiency, and decision-making tasks, consistently outperforming top retrieval-based memory approaches with more efficient token usage. EvoLib also shows robust performance across random task orders, indicating practical applicability in real-world scenarios where tasks arrive arbitrarily. The system achieves higher performance gains with increasing test-time compute compared to isolated compute scaling and other memory-based methods. Code and experiments are publicly available on GitHub.