Microsoft's EvoLib lets AI learn from experience in real-time
EvoLib turns raw AI experience into reusable knowledge—without model updates
Microsoft Research has unveiled **EvoLib**, a self-supervised framework designed to let large language models (LLMs) learn from their own experiences during inference—without requiring external labels or model updates. Unlike traditional memory systems that store raw experiences as static archives, EvoLib transforms past attempts into reusable skills and reflective insights, continually refining and reweighting them to generalize across tasks.
The system introduces two core mechanisms: **consolidation** and **dynamic weighting**. Consolidation merges new knowledge with existing insights to create more general solutions, while the weighting mechanism prioritizes knowledge based on its long-term utility. Tested on tasks like mathematical reasoning, code generation under constraints, and long-horizon decision-making, EvoLib outperformed top retrieval-based memory approaches while using fewer tokens. Crucially, it works with any black-box model via APIs, making it deployable without retraining the underlying AI.
- EvoLib enables **self-supervised learning** during inference, extracting reusable skills and insights from raw experiences without external feedback.
- It uses **consolidation** and **dynamic weighting** to refine knowledge over time, improving performance across tasks like math, coding, and decision-making.
- The framework works with **any black-box model** via APIs, requiring no model updates—making it practical for existing deployments.
Why It Matters
EvoLib bridges the gap between memory and learning, enabling AI agents to improve continuously without retraining—unlocking smarter, more adaptive systems.