What does EvoLib actually add to a deployed agent?

EvoLib gives the agent a persistent library of modular skills and short reflective insights that it writes for itself from its own inference trajectories. Similar entries are merged into more general forms and each entry is re-weighted by long-term usefulness, so the library keeps improving between runs without any fine-tuning of the base model.

How is EvoLib different from a standard retrieval-augmented memory?

Standard RAG memories only grow: every trajectory gets stored and retrieved by similarity, and stale or misleading entries dilute quality. EvoLib actively consolidates duplicates into more general skills and downweights entries whose long-term utility falls, closer to an evolving library than a log — which is why Microsoft calls it test-time LEARNING, not test-time memory.

What tasks did Microsoft evaluate EvoLib on?

The EvoLib paper reports gains across math reasoning, code generation under strict efficiency constraints, and long-horizon agentic tasks, beating top retrieval-memory baselines and test-time-scaling baselines across most of the compute range. Curves are qualitative in the blog; the arXiv preprint (2605.14477) has the full tables and ablations.

How do I try EvoLib today?

Clone github.com/microsoft/EvoLib — the MIT-licensed code implements the evolving library on top of a standard LLM inference loop. Microsoft also ships EvoLib in Azure AI Foundry Labs for teams that want a managed environment; the underlying method is open so on-prem or self-hosted deployments are straightforward.

Microsoft Research · 2026-07-30 · major

Microsoft EvoLib — test-time learning turns agent trajectories into an evolving skill library

EvoLib gives a deployed agent a growing library of skills and reflective insights, built only from its own runs. Similar entries are consolidated, low-utility ones fade, and the base model is never fine-tuned — published as an MIT repo and an arXiv paper.

Microsoft Research EvoLib blog banner — turning experience into evolving knowledge

EvoLib turns every inference trajectory into a library entry, then consolidates duplicates and prunes low-utility skills — no fine-tuning.

Quick facts

Maker	Microsoft Research (Redmond, Montréal, Health Futures)
Artifact	Test-time learning framework + open repo
Code	github.com/microsoft/EvoLib (MIT)
Paper	arXiv:2605.14477 — 'Test-Time Learning with an Evolving Library'
Domains tested	Math reasoning, code under efficiency constraints, long-horizon agent tasks
Base model change	None — no gradient updates, no external supervision
Access	Ships behind Azure AI Foundry Labs — open code + arXiv

What is it?

EvoLib is Microsoft Research's test-time learning framework. A deployed agent writes short skills and reflective notes from its own runs into a persistent library, and the library itself is updated between calls — similar entries merge, low-utility entries fade, and the base LLM is never touched. Code is MIT on GitHub.

How does it work?

Two operators run alongside inference. A consolidator finds near-duplicate skills or notes and rewrites them as one more general entry, and a utility re-weighter tracks how much each entry actually improves later trajectories, decaying entries that stop being useful. The result is a library that grows in coverage while shrinking in redundancy.

Why does it matter?

Fine-tuning a frontier model to add a skill is expensive and often regresses others. EvoLib gives a deployed agent a way to keep learning on the job — math reasoning, efficient code and long-horizon planning — while beating retrieval-memory and test-time-scaling baselines. Enterprises can ship one base model and let each deployment specialise itself.

Who is it for?

agent teams, applied research groups, enterprise LLM ops

Frequently asked questions

What does EvoLib actually add to a deployed agent?: EvoLib gives the agent a persistent library of modular skills and short reflective insights that it writes for itself from its own inference trajectories. Similar entries are merged into more general forms and each entry is re-weighted by long-term usefulness, so the library keeps improving between runs without any fine-tuning of the base model.
How is EvoLib different from a standard retrieval-augmented memory?: Standard RAG memories only grow: every trajectory gets stored and retrieved by similarity, and stale or misleading entries dilute quality. EvoLib actively consolidates duplicates into more general skills and downweights entries whose long-term utility falls, closer to an evolving library than a log — which is why Microsoft calls it test-time LEARNING, not test-time memory.
What tasks did Microsoft evaluate EvoLib on?: The EvoLib paper reports gains across math reasoning, code generation under strict efficiency constraints, and long-horizon agentic tasks, beating top retrieval-memory baselines and test-time-scaling baselines across most of the compute range. Curves are qualitative in the blog; the arXiv preprint (2605.14477) has the full tables and ablations.
How do I try EvoLib today?: Clone github.com/microsoft/EvoLib — the MIT-licensed code implements the evolving library on top of a standard LLM inference loop. Microsoft also ships EvoLib in Azure AI Foundry Labs for teams that want a managed environment; the underlying method is open so on-prem or self-hosted deployments are straightforward.

Try it

git clone github.com/microsoft/EvoLib