A model has no memory of yesterday and no access to your files, unless your tool has been given them. ChatGPT, Claude and Gemini all ship some form of memory across conversations now, and file connectors are common. Where those are switched on, whatever they pull in arrives as more text in the same window. That is the point: the model has exactly one thing, the text sitting in front of it right now.
Everything it appears to know about your situation, it knows because something put it there. That is why the fix for a bad answer is so often "give it the actual material" rather than "word the question more cleverly."
It is also why more context stops helping much sooner than most people expect.
What the window actually holds
The context window is the total amount of text the model can consider at once. Everything competes for the same space:
- Any standing instructions the tool sends on your behalf
- Anything it has remembered about you from earlier conversations
- Whatever you have pasted or attached
- Every message in the conversation so far
- Every reply the model has already given
That last one surprises people. A long back-and-forth fills the window with the model's own output, not just yours.
The important part is what happens each time you hit enter: the whole lot is re-read from scratch. The model is not remembering your earlier messages. It is re-reading them. There is no accumulated understanding carried between turns, only a transcript that gets longer.
So a context window is a desk, not a filing cabinet. Everything you want considered has to be on the desk. And a desk can be too crowded to work at.
Why more context makes answers worse
Past a point, adding material stops helping and starts hurting. Three distinct things go wrong, and they are worth being able to tell apart.
Attention spreads thin. The more you give it, the more competition your one important sentence has. Material in the middle of a long input fares worst: the beginning and the end carry more weight, often enough that it is worth planning around, though how much varies by model. Bury your one crucial sentence in the middle of forty pages and it will be treated as one sentence among thousands, because that is what it is.
Contradictions get resolved by guessing. Paste a folder of documents and some of them disagree: an outdated spec next to the current one, an old price list, a decision that was later reversed. You know which is authoritative. The model does not. It sees two claims with equal standing and picks one.
Irrelevant material invites irrelevant answers. Give a model forty pages, and it will behave as though those forty pages are relevant, because you presented them as relevant. Ask a narrow question against a broad document and you often get a broad answer that dutifully touches everything you supplied.
None of these are bugs that get patched. They follow from how the thing works, so the skill is choosing what goes on the desk.
The test for what to include
Before adding something, ask:
Would a competent freelancer need this to do the task?
Someone capable, dropped into your situation this morning, with no access to your systems and no history with your team. What would they need you to hand them?
That test cuts both ways, and the second half matters more:
Include what only you know. The constraint that is not written down anywhere. The thing you already tried that did not work. Who the output is for. The format it has to land in. This is the material people leave out most often, and it is the material with the highest return.
Leave out what is ambient. General knowledge about your industry. Standard definitions. Background that explains the situation to a human but does not change what a good answer looks like. If removing a paragraph would not change the correct answer, it is costing you attention and buying nothing.
A structure that works
For anything with real material attached, this order holds up:
- The task, in one or two sentences. What you want, and who it is for.
- The material. The document, the code, the transcript, the data.
- The ask, restated. Specific, and pointing at the material.
Putting the task first means the model reads the material already knowing what it is looking for, the way you would skim a contract differently if you had been told to check the termination clause.
When there is genuinely too much
Sometimes the material really is enormous. The instinct is to paste it all and hope. Two passes beat one.
First pass, extract. "From the document below, pull out every clause that mentions payment terms. Quote them exactly, with the section number. Do not summarize or interpret."
Second pass, work. Start fresh, paste only the extract, and ask your real question against that.
You have swapped forty pages of mixed relevance for one page of directly relevant quotes. The second answer is working from a desk that has only the right things on it. The extraction step is also checkable in a way a single mega-prompt never is. You can read the quotes and confirm nothing important was missed before you build anything on them.
Your conversation is context too
This is the one people miss. A thread that has run for thirty messages is carrying all thirty, including:
- The approach you tried for ten messages and abandoned
- The misunderstanding you spent three messages correcting
- Two versions of a draft you have since rejected
All of it is on the desk, all of it is being re-read, and none of it is labeled as dead. Worse, a model that has committed to a position tends to stay consistent with it, including consistently wrong.
Signs you have overfilled it
- The answer restates your input back at you instead of doing the task
- It touches every topic you supplied, weighting them all equally
- It uses a detail you included as background as though it were the point
- It contradicts something in the middle of your material
- Quality got worse when you added something, not better
That last one is the clearest signal, and it is worth trusting when you see it. The instinct on a bad answer is to add more. Try removing instead.
The short version
The model only knows what is in front of it, so give it the things only you know, and nothing else. Put the task before the material and repeat it after. When the material is too big, extract first and work second. When a thread has wandered, start a new one.
Context is a selection problem, and you are the only one who can do the selecting.
The companion to this is How to Write Better Prompts, which covers the shape of the ask itself. This post is about what you put behind it.