Watch on YouTubeWatch on Instagram
About this video
A million tokens at once, and AI still forgets you. Here's why.
What a context window really is, why the model re-reads your whole chat every message, and a real test of what gets lost inside it.
Episode 5 of Learn AI in Under 3 Minutes.
Measured for this video: one made-up sentence hidden at the start, middle or end of The Adventures of Sherlock Holmes (Project Gutenberg #1661), cut to 1,000, 8,000, 32,000 and 100,000 tokens, then asked to Qwen3.5-0.8B (open weights, February 2026; greedy decoding, thinking off) with a question in the sentence's own words or reworded. 3 hidden sentences x 3 spots per length. Same words: found 35 of 36 times (89% at 100,000 tokens). Reworded: 89%, 100%, 67%, 44% at the four lengths; hidden in the middle of the 32,000 and 100,000-token books, missed 6 of 6, answering with other things from the book ("In the bank cellar", "Portugal"). One small model is an example, not a benchmark: big models do much better.
Context windows: OpenAI, Anthropic and Google model docs, October 2026 (GPT-6: 1.05M tokens; Claude 5.5 and Gemini 4 Argon: 1M). Price: Anthropic pricing, Claude Opus 5.5 at $4 per million input tokens, cache reads $0.20. Chroma, "Context Rot" (July 2025, 18 models). OpenAI Memory FAQ: saved memories are part of the context ChatGPT uses to generate a response.
0:00 A million tokens
0:16 The desk
1:09 My test
1:43 Memory
Music: Special Spotlight by Kevin MacLeod, CC BY 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Special_Spotlight_(ISRC_USUAN1600067).mp3)
Follow @PranavInnovates for the tech story that matters, every day.
What the video says
I hid one sentence in a Sherlock Holmes book, then asked an AI to find it. Where I hid it changed everything. GPT-6, Claude and Gemini read about a million tokens at once: this whole book, seven times over. So do they remember you? Not even a little. That million is the context window: the model's desk. Its instructions, your messages, its replies, your files, all on one desk. Off the desk? Invisible. And here's the surprise: it keeps nothing between messages. Every time you hit send, the app puts the whole chat back on the desk, and the model reads it all again. Quick one. Your chat is a hundred messages long. To answer the next one, what does it read? Just your new message, or all hundred? All hundred. Every time. Fill Claude Opus 5.5's desk, and each reply costs four dollars. That's why episode three's prompt cache matters: re-reads drop to twenty cents. The desk also has an edge. When it's full, apps drop the oldest messages, or squash them into a summary. That thing you said on day one? Gone. But the real catch is inside the window. I tested it. I hid one line in the book, then asked an open model, Qwen 3.5, to find it. Same words in the question? It found it almost every time, even at a hundred thousand tokens. Reword the question, and it dropped to forty-four percent. In the long book, with the line in the middle, it missed every single time, and grabbed something else from the book: the bank cellar. Portugal. Big models do far better, but they slip too. Chroma tested eighteen of them: the longer the input, the less reliable they got. So where does ChatGPT's memory feature actually live? Inside the model, or somewhere else? Somewhere else. OpenAI says saved memories are part of the context. It's a note, put back on the desk, in every chat. So a context window is a desk, not a brain. A bigger desk doesn't remember. What you put on it decides the answer. Here's how I build. I treat the window like RAM, not storage. Only what this step needs goes on the desk. Everything else lives in files the agent can look up. Choosing what goes on the desk is a skill: episode seven. Next: how AI turns static into pictures.