Microsoft introduced EvoLib to help AI models learn from experience
The new framework enables AI agents to refine and reuse knowledge over time without requiring updates to the underlying models.
Microsoft researchers published a paper detailing EvoLib, a framework that allows language models to turn their past interactions into reusable knowledge. Instead of just storing every raw conversation in an archive, the system extracts useful skills and insights. It then refines this information so it can be applied to new tasks later.
The framework improves performance by consolidating knowledge and weighting it based on its future utility. This process happens during inference, meaning the system does not need to undergo additional training or parameter updates to learn. It functions with black-box models accessed through APIs.
The key to better learning may not simply be storing more memories or spending more compute.
In testing, the researchers evaluated the method across three domains: mathematical reasoning, coding efficiency, and long-horizon decision making. The approach outperformed traditional memory-based methods in these benchmarks while using fewer tokens. It also showed resilience when task orders were shuffled.
Despite these results, the system depends on the quality of the initial experiences provided to the model. The research does not test how the library scales across extremely long durations or significantly larger heterogeneous datasets, nor does it replace the need for base model improvements.
Source: Microsoft Research Blog