EvoLib, explained: a Microsoft Research framework for AI that learns on the job

The "memory" features shipping in today's AI products are mostly a growing pile of past prompts and answers. Microsoft's new research framework, EvoLib, names the missing step: turning each inference trajectory into reusable knowledge the model can lean on the next time it faces a similar problem, all without retraining and without any ground-truth labels.

EvoLib is the work of Microsoft Research and is described in an accompanying arXiv preprint (2605.14477, "Test-Time Learning with an Evolving Library"), with code released on GitHub. The Microsoft Research blog post frames it as a position on a problem the field has been circling: a raw archive of experience is not the same as a learned lesson.

The framework sits in the slot researchers call test-time learning, the moment a model is being used. Instead of a separate training run, EvoLib treats each inference trajectory, the chain of reasoning, code, or agent steps the model produced for a given task, as raw material. The system extracts two kinds of abstractions from that material: modular skills, which are reusable procedures, and reflective insights, which are higher-level lessons distilled from several trajectories. Both go into a shared library that subsequent inferences can draw on.

The library is not just a flat cache. EvoLib applies a principled weighting and consolidation mechanism that trades off immediate utility, how much a particular abstraction helped on the problem it came from, against long-term value, how general the abstraction is and how often it gets reused. Simple, instance-specific lessons can be promoted into more general reusable skills; redundant or low-value entries get pruned. That progression, from specific to general, is the "evolving" in the name.

Because no model parameters are touched, EvoLib can be applied to any language model or AI system served through an API, including closed-weight frontier models the user cannot see inside. In practice, a research lab could in principle wrap a third-party model with EvoLib and let it accumulate experience across runs, without waiting on the model vendor to retrain.

The reported evaluation surface is mathematical reasoning, code generation, and multi-turn agentic environments, where the authors say EvoLib beats top test-time scaling and learning methods that also avoid ground-truth feedback. The paper's full ablation tables and baseline definitions live behind the arXiv HTML, which in the public mirror is truncated past the abstract, so the specific benchmark numbers and the lineup of "top" baselines are not independently verifiable from the materials in hand. The blog post, paper, and code are also a single-vendor bundle: there is no independent reproduction or third-party benchmark in the source set, and EvoLib is a research framework, not a shipped Microsoft product feature.

What the framework actually changes depends on what kinds of abstractions a model can reliably extract from its own work. The library is only as good as the extraction step and the weighting step, and those are still author-reported at this point. The next read-throughs to wait for are full benchmark numbers in the arXiv HTML once it is no longer truncated, an independent reproduction, and any third-party measurement on closed-weight models wrapped with the library.