Microsoft Research's EvoLib Turns Inference-Time Experience Into Reusable Skills Without Ground-Truth Labels

Microsoft Research·high signalXBlueskyLinkedInCopy link

Microsoft Research published EvoLib on July 30, a framework that transforms a model's past attempts into consolidated reusable skills and reflective insights rather than accumulating raw memories, with similar knowledge units merged into generalizable forms and importance dynamically reweighted by long-term utility. Evaluated on math reasoning, code under efficiency constraints, and long-horizon environment decisions, it beats top retrieval-based memory and other abstract memory mechanisms at lower token usage, with gains widening as test-time compute increases. It targets black-box API-deployed models and code is on GitHub — directly applicable to anyone running a persistent agent loop.