Microsoft unveils EvoLib for continuous LLM learning

Microsoft Research has developed EvoLib, a novel framework designed to help large language models learn from their own execution history at runtime. Detailed in a paper titled "Test-Time Learning with an Evolving Library," the system operates without requiring ground-truth labels, external feedback, or updates to the underlying model weights. This self-supervised approach allows EvoLib to be applied directly to black-box language models and commercial AI systems deployed through standard APIs.

Instead of treating memory as a static, ever-growing archive of raw interaction histories, EvoLib extracts reusable skills and reflective insights from past successes and failures. The framework continually refines this knowledge through two primary mechanisms: consolidation and dynamic weighting. Consolidation merges newly extracted insights with similar existing library entries to create more generalized, transferable skills. Meanwhile, the weighting mechanism dynamically updates the importance of each knowledge unit based on its long-term utility and its contribution to solving future tasks.

Researchers evaluated EvoLib across three distinct benchmarks, including tasks that required writing code under efficiency constraints and making decisions to navigate long-horizon environments. The framework consistently outperformed top retrieval-based memory methods and other abstract memory mechanisms while maintaining more efficient token usage. Furthermore, EvoLib proved highly robust to the order in which tasks were encountered, showing stable performance even when handling a mixed stream of heterogeneous requests. The code and experimental results have been made publicly available on GitHub.

For AI practitioners, EvoLib shifts the paradigm of agent memory from simple retrieval to active, evolving synthesis. Rather than bloating context windows with raw historical logs that degrade performance and spike API costs, developers can deploy agents that distill their own best practices over time. This allows deployed systems to become progressively smarter and more cost-effective the more they are used, without the need for expensive fine-tuning runs or structured training curricula.