EvoLib от Microsoft Research: языковые модели учатся на опыте без переобучения

Microsoft Research has announced EvoLib — a framework that transforms the accumulated experience of language models into reusable skills and knowledge, helping LLMs adapt to new tasks months after deployment without additional training.

Why simple memory isn't enough

EvoLib is built around a thesis that the Microsoft Research team states directly: language models don't get smarter just by remembering more.

"LLMs do not get smarter just by remembering more", — stated in the official

EvoLib description from Microsoft Research.

Expanding the context window and archiving dialogues is storage, not learning. If a model has processed thousands of code analysis or documentation writing tasks, it should derive principles and patterns — not simply accumulate examples.

Modern LLMs have several layers of "memory": the context window (temporary), external knowledge bases like RAG (searching for similar fragments), and parametric memory (model weights). EvoLib adds a fourth layer — a dynamic skill library that grows from experience and is refined during real-world operation.

After deployment, most LLMs effectively freeze: tasks and context change, but the model stays the same. Traditional adaptation via fine-tuning requires data preparation, computational resources, and often temporary system downtime.

How the evolving knowledge library works

EvoLib extracts reusable skills and insights from each interaction, forming an "evolving knowledge library." This is not a cache of responses or a dialogue log — but structured abstractions: what approach worked when solving a task, what proved effective.

The key difference from standard LLM memory approaches lies in the nature of what is stored. A RAG system finds a similar past response; EvoLib extracts a generalized skill applicable to new situations. When encountering an unfamiliar task, the model turns to this arsenal and applies patterns from past experience — while the library continues to grow and refine.

In the context of AI agents, this is particularly valuable: an agent that has solved hundreds of similar tasks in a corporate environment should become more effective — better at choosing tools, avoiding past mistakes, applying optimal action chains. EvoLib was designed precisely for such long-term scenarios.

The name EvoLib (from Evolving Library) reflects the core idea: knowledge is not static — it evolves together with the model's practice. According to Microsoft Research's concept, this approach is conceptually close to how humans learn — through the accumulation of generalized experience, not through rote memorization of examples.

What this means for AI in production

EvoLib offers an alternative to costly retraining cycles. The problem of quality degradation after deployment is well known in the industry: data shifts, tasks become more complex, and the model gradually falls behind reality. If EvoLib proves scalable, it changes the logic of AI system lifecycles: instead of periodic fine-tuning — continuous knowledge accumulation directly during operation.

For enterprise AI systems that process repetitive tasks for months — customer support, document analysis, process automation — organic adaptation without system downtime represents tangible practical value.

Microsoft Research has so far described EvoLib at a conceptual level: specific benchmarks and implementation details are not disclosed in the announcement. Nevertheless, EvoLib addresses one of the fundamental open problems in applied AI — how to make a model get smarter from working, not just from training.