Context windows
- foundations
- long input
A context window is the total amount of text a model can consider at one time. Your current message, earlier turns, pasted documents and the model’s own replies all share it. When people say a model “remembers” — this is what they mean: the text still inside the window.
No window, no memory
Outside the window there is no memory to search, because there is no store: the model reads what it is given, each time. A new conversation starts with an empty window — which is why a chat that “knows” your project today knows nothing about it tomorrow unless you bring the context along.
What happens at the limit
Two things, and both matter:
- Truncation. Systems cut from the oldest end first — measured in tokens, the start of a long chat is the first to fall out. Some tools warn you; many do not.
- Attention dilution. Even content still inside a very large window competes for attention. Instructions buried deep in a 100-page paste are easier to lose than the same instruction stated right where it is needed.
Watch the window fill — and start pushing the oldest text out:
The context window everything the model can see right now
New messages push in from the right — and the oldest text slides out of the window on the left. Outside the window there is nothing: the model cannot look it up later.
A bad example
Three hours into a working session you ask: “Based on what I told you at the start, which vendor did I prefer?”
The first hour may have fallen out of the window already. The model answers from what remains — and rather than say “that part is gone”, it may guess. Fluently.
A better example
Ask: “Summarize the decisions we’ve made so far as bullets. Then I’ll start a fresh chat and paste that.”
A fresh window full of relevant, compact context beats a stale window full of transcript. This carry-forward habit is the single most useful long-session tactic.
Why it works
Treat the window as working memory you manage:
- Keep what is relevant. Summarize and carry forward; drop the rest deliberately.
- Restate what matters. If a rule is critical, say it near the work it governs — not three hours earlier.
- Start fresh when it gets muddy. A new chat with a good brief beats a long one with a bad memory.
A habit that keeps long sessions useful: whenever the goal shifts, restate the standing facts in the new message — the reader, the constraints, the format. Three lines of recap cost seconds and protect everything that came before; drift happens exactly where a recap was skipped. The model never gets tired of instructions — it only stops seeing them.
Practice
Fix the forgetting chat
You have been working in one long chat for hours. You ask about a preference you stated near the beginning, and the answer is confidently wrong — the start of the conversation has fallen out of the context window.
What is the best next step?
Hint
What is the smallest, most relevant context you could bring to a clean window?
Why this is the answer
The window is finite working memory, and the oldest content falls out first. The durable fix is a carry-forward habit: summarize the essentials, start fresh, paste the brief. Restating a line helps briefly; managing the window deliberately helps permanently.
Transfer
- Long drafting project: keep a running “brief” to paste at the top of new sessions.
- Research with many sources: work one source at a time, extract notes, then combine.
- Any answer that contradicts something you said early: suspect the window before the model.
Key takeaways
- The context window is everything the model can see at once — your message, the history, the documents.
- At the limit the oldest content is dropped first, and instructions buried in a huge paste compete for attention.
- Carry a short brief forward into fresh chats instead of hoarding transcripts.
Next
Next: Why AI makes things up — the failure mode both mechanisms predict. The context window glossary entry covers the basics.