context stuffingcontext stuffing
Context stuffing is the practice of loading relevant information into an link not tracked's context window before asking it to perform a task, so the model can reason over material it wouldn't otherwise have access to.
Rather than relying solely on what the model learned during link not tracked, you deliberately "stuff" the context with documents, code, conversation history, or other reference material that the model needs right now.
This is analogous to a person gathering all the relevant papers onto their desk, and perhaps working with spaced repetition to cache the ideas in their head before sitting down to work through a problem — it's not about permanently memorizing the material, but about having it available in link not tracked for the duration of the task.