Why AI “Forgets” What You Said Ten Minutes Ago
You’ve probably had this happen: you give an AI tool a detailed instruction early in a conversation, keep chatting for a while, and then it suddenly acts like that instruction never existed. It’s not being lazy or inconsistent — it’s running into its context window, the fixed amount of text the model can actually “see” at once when generating a response. Understanding what that window is, and how it actually empties out, changes how you prompt.
What a Context Window Really Is
Every message you send, every reply the AI gives, and every file you paste in gets converted into tokens — small chunks of text, roughly three-quarters of a word each — and all of those tokens get fed back into the model together every single time it generates a new response. The context window is the maximum number of tokens that can fit into that single pass. It’s not a memory in the human sense; the model isn’t “recalling” your earlier message, it’s re-reading the entire conversation from scratch each time, the way you’d re-read a whole document before writing one more sentence in it. Once the conversation grows past the window’s limit, the oldest content gets trimmed to make room, exactly like a word processor scrolling the top of a document out of view.
Where This Quietly Breaks Your Results
This is why a long conversation gradually drifts away from instructions you gave at the start: your “please keep responses under 100 words” from an hour ago may simply no longer be in the window the model is reading from. It’s also why pasting a huge document and then asking a detailed follow-up question sometimes gets a vague answer — the model’s attention is spread thin across everything in the window, and details buried in the middle of a long input get weighted less reliably than details near the beginning or end.
Three Habits That Work With the Window Instead of Against It
First, restate critical instructions periodically in long conversations rather than assuming they persist — a quick “reminder: keep this under 100 words” costs you a sentence and saves you a rewrite. Second, put the most important information at the start or end of a long prompt, not buried in the middle, since that’s where models are most reliably accurate. Third, when a conversation has clearly drifted or started producing inconsistent answers, don’t keep patching it — start a fresh conversation with a tighter, more complete prompt. A clean context window with the right information in it consistently outperforms a long, cluttered one trying to hold everything you’ve ever said.