The paper introduces skill internalization as a process that converts externally specified procedural skills into persistent capabilities of language-model agents. The goal is to let agents execute these skills at inference time without needing to retrieve or reinject skill descriptions, which would normally require extra context or prompting.

The authors identify a failure mode where skills remain in the model but access fails—meaning the internalized knowledge is present yet not reliably available when needed. To address this, they propose factorized skill memory, a design that separates how skills are stored from how they are accessed, aiming to make continual internalization more robust.

Because the abstract is brief, the specifics of the factorization and evaluation are not detailed here. The work appears to target a practical bottleneck in deploying language agents that must accumulate and use many procedures over time.