Contextual knowledge

Contextual knowledge is what the model knows from the current request: the files the agent has read, the docs you pasted, the error output, the instructions in your project file. It is the counterpart to parametric knowledge. Anything in the context window can override what the model would otherwise assume, which is the whole reason coding agents read files before editing them instead of writing from memory.

It is also the more reliable of the two, as long as it is actually there and still accurate. Models generally trust the context over their training when the two conflict, so a pasted changelog beats a stale habit. The failures come when people assume something is in context and it is not (the agent never opened that file), or when it was in context but has since changed on disk (you edited it after the read).

Make contextual knowledge deliberate. Point the agent at the specific files that define the truth for this task, refresh them after big changes, and prefer primary sources over your own summary of them. When an answer seems to ignore your codebase, ask what it read; the honest reply is often 'nothing yet'.