EvoLib: Turning experience into evolving knowledge
| Source: Microsoft Research Blog
Tags: Microsoft Research, EvoLib, test-time learning, LLM agents, continual learning, AI memory
Microsoft Research's EvoLib enables language models to learn from their own inference-time experience without retraining—turning past successes and failures into a self-refining library of reusable skills and insights that improve future task performance.
Details
EvoLib, introduced in the Microsoft Research paper 'Test-Time Learning with an Evolving Library,' provides a framework for LLMs to improve continuously during deployment without model updates. Instead of accumulating raw experience as a growing archive, EvoLib extracts reusable skills from successful solutions and reflective insights from failures, then continually refines and consolidates this knowledge as new experiences arrive. The mechanism is self-supervised—no ground-truth labels or external feedback are required. The evolving library applies to any black-box language model or AI system accessible via API, since it doesn't require access to model weights. Over time, instance-specific observations generalize into broader transferable knowledge for novel tasks. This addresses a known limitation of standard memory-augmented LLMs: raw experience accumulation becomes a retrieval problem, not a learning problem. EvoLib reframes this by treating the library as a continuously evolving knowledge base rather than a static archive. The approach is particularly relevant for agentic deployments where models complete similar task types repeatedly over long operational horizons—the model gets measurably better at those tasks without any retraining.