SkillSmith is a research system for composing two kinds of agent memory that are usually kept apart: text-based knowledge from prior experience and parametric skill libraries in model-weight space.
The paper instantiates parametric learning with prefix-tuning and augments an LLM so it can ingest both prefix weights and textual context. The output is not a prompt or a tool call; it is a new set of prefix weights intended to express a target skill.
That makes the story relevant beyond fine-tuning. Agent systems already have files, tools, memories, and recipes. SkillSmith points at another artifact type: small learned skill modules that can be selected, described, combined, and generated as part of an agent workflow. If this line works, skill registries stop being only prompt libraries and start looking like mixed text-and-weight catalogs.