Lesson 3 of 6
Vague memory vs fresh notes
5 min read
Two ways to ask the same question. Ask a chatbot from cold memory what your company's refund policy is, and you get a confident maybe. Paste the policy in first, and you get the exact rule — with the line it came from.
Its memory is a blur, not a filing cabinet
Everything a model 'knows' is squeezed into its internal settings during training — the gist survives, the fine print blurs. So from memory it's great on the shape of a topic and shaky on the exact number, date, or clause. It doesn't store your documents at all; at best it half-remembers something similar it once read.
Asking from memory is like asking a well-read friend to quote a contract from a book they skimmed years ago. They'll get the gist. They'll miss the specifics.
Put the source in front of it
When you paste the actual text — the policy, the email, the page — into the chat, the model no longer has to recall anything. It reads the facts right there and answers from them. This is called grounding: the answer is anchored to a source you provided, not dredged from fuzzy memory, and it can point to exactly where it got each part.
Same model, same question — but a grounded answer is more accurate and checkable, because you can see the source it leaned on.
Whenever accuracy matters, don't make it remember — make it read. Paste in the document, the numbers, the email thread, and ask it to answer only from what you gave it.
What to take away
- —From memory, a model is good on the gist and unreliable on the specifics.
- —Give it the source in the chat and it reads instead of recalls — far more accurate.
- —A grounded answer can cite where it came from, so you can check it in seconds.
You need the exact cancellation fee from a contract you have as a PDF. What gets the most reliable answer?
Continue in the app
Take the whole When AI Gets It Wrong course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.