Parametric knowledge
Also called Baked-in knowledge · Training knowledge
What the model knows because it was in the training data, stored in its weights and frozen from that moment on.
Parametric knowledge is everything a model knows without being told in the prompt. It was absorbed during training and lives in the parameters: how Python works, what React looked like at the time, common bugs, idioms, the names of popular libraries and their older APIs. It is why an agent can write a working Express server from a one-line request with no documentation loaded.
The catch is that it is a snapshot. Nothing after the knowledge cutoff is in there, and even things before it are compressed and approximate. When your project uses a framework version the model never saw, or a private internal library, its parametric knowledge does not go blank; it fills the gap with the nearest thing it does know, which is how you get last year's API under this year's package name.
Know which kind of fact you are relying on. For stable, widely used things, parametric knowledge is fast and usually right. For anything recent, niche or specific to your codebase, put the real material in the context window so the model works from contextual knowledge instead. A good rule: if you would need to look it up, so does the model.
- Most toolsNothing in the harness updates what the model knows; docs lookups and web fetch are how fresh facts get into a request.
- Claude CodeA CLAUDE.md note like 'we use Tailwind v4 syntax, see docs/tailwind.md' steers the agent off stale parametric habits.
“How does it know the whole Next.js router API without me pasting anything?”
“Parametric knowledge, from training. Just check which version it's remembering before you trust it.”
Course 01 puts every one of these terms to work: you install Claude Code, run the loop, and ship a real project — permission modes, compaction, hooks and all.
Start Course 01 →