Context Window, Memory and Projects: What the AI Actually Sees

The context window: the AI's desk

Picture a desk. Everything the AI can read before answering has to fit on it: your messages, its own replies, any files you attached, your project instructions, whatever its memory has saved about you. That desk is the context window. Gemini's help page describes it as the AI's "reading capacity".

If something isn't on the desk, the AI doesn't know it exists. Yesterday's conversation, the comment your client made that you never typed in: none of it reaches the model unless a feature like memory or projects puts it there for you.

The size of the desk is measured in tokens, the units a model splits text into. For a sense of scale, Anthropic says 200,000 tokens is about 500 pages of text.

How much fits? ChatGPT, Claude and Gemini

As of October 28, 2026, the three companies list very different sizes, and the numbers change from one month to the next.

• ChatGPT: according to the pricing page, the free plan gets 27K tokens in Instant mode, and you can only send it about 12 pages of text. Go and Plus get 54K in Instant and 256K in reasoning mode, roughly 320 pages. Pro goes up to 400K.
• Claude: on paid plans, recent models such as Opus 5.5 and Sonnet 5.5 have a one-million-token window, five times the 200K of the models with the smallest one. The page gives no figure for the free plan.
• Gemini: 32K tokens without a subscription, 128K with Google AI Plus, one million with AI Pro and Ultra.

On a free plan, a long document may not fit: those 40 pages are well over the 12 pages ChatGPT accepts in Instant mode. And you don't get the whole desk. OpenAI notes that the room left for your input is smaller than the total window, because system instructions, memory and the model's own reasoning take up part of it.

More isn't better: why the AI loses track

You might expect a bigger desk to solve everything. Anthropic says otherwise: in its documentation, more context isn't automatically better. Its engineers found that as the window fills up, the model gets worse at recalling what's in it accurately. They describe an "attention budget" that every addition draws on.

So Anthropic recommends finding the smallest set of useful information that gets the best result, and calls this work context engineering.

Give it the right document

If your question is about chapter 3 of your thesis, paste chapter 3 and the outline. Leave the other 80 pages out. If you attach a long file, point to the spot: "pricing is on page 12." A consultant preparing for a client call pastes the notes from the last meeting and the proposal that went out; the full email thread adds nothing. The AI then answers from your source, which is what people mean by grounding.

Tell it what's missing

OpenAI advises you to provide enough context for the model to understand the request. Google's Gemini prompting guide goes further and says to provide as much context as possible. So the three companies don't say the same thing. Our take: follow Anthropic's rule for long documents and Google's for three-line requests, where the real risk is saying too little. When in doubt, ask the AI what it's missing.

For the rest of your request, our anatomy of a great prompt walks through each part.

Message, project or memory: where to put what

ChatGPT, Claude and Gemini each give you several places to store information. They don't last the same amount of time.

The message, for a one-off task

Whatever you type or attach in a conversation only serves that conversation. It's the right place for today's document or an instruction you won't need again.

The project, for recurring work

A project groups conversations, files and shared instructions. In Claude, each project has its own chat history and knowledge base, and free users can create up to five. In ChatGPT, a project holds 5 files on Free, 25 on Go or Plus and 40 on Pro, with an option for project-only memory. The Gemini app added notebooks in April 2026. And NotebookLM is now called Gemini Notebook.

A teacher could set up a "10th-grade history" project with the syllabus and grading rubric, then open one conversation per unit. Anthropic recommends keeping project instructions concise and focused on what's essential. This prompt gives you a first draft:

Memory, for what's true everywhere

Memory carries information from one conversation to the next: your job, your preferences, an ongoing piece of work. In Claude, it is on by default for Free, Pro and Max, and each project has its own. It's the same memory in chat and in Cowork, and you can edit every entry. Searching past chats, though, requires a paid plan.

ChatGPT draws on saved memories and your chat history, and the features vary by plan, region and platform. Gemini's memory is limited to users 18 and over on a personal account with Activity turned on, and it doesn't work in Live chats.

To sort things, ask how long the information stays true. Your job goes into memory, this year's syllabus into a project, and today's quote stays in the message. Check what memory has saved from time to time, too: an outdated detail like "I'm a sophomore" two years after graduation keeps steering the answers. We prefer a short memory: five true lines beat a long profile that's half out of date. For Claude Cowork, our guide on giving Claude Cowork a memory covers the settings. Our guides to prompting Claude and Gemini go tool by tool.

When a conversation gets too long

A two-hour thread eventually clutters the desk. When code execution is turned on, Claude summarizes earlier messages as the window fills. A summary loses detail. Be suspicious when the AI reopens a decision you already made or asks for something you already told it.

For agents that run for a long time, Anthropic lists three techniques: summarizing, taking notes and handing subtasks to other agents. The first two work fine by hand. Ask for a summary, start a new conversation and paste it in. You're back to a clean desk.

The problem gets worse when the AI works on its own. An agent chains dozens of actions, and every page it reads lands on the desk. Our article on AI agents explains what sets an agent apart from a chat, and our article on AI loops covers how to make it repeat a task until a stop condition is met.

Gems are going away: what to do with your instructions

If you keep instructions in Gems on Gemini, mark two dates. For personal accounts, Gems become read-only on November 17, 2026 and are removed on December 17. Google is replacing them with skills, which are reusable custom instructions.

Before then, copy the text of each Gem into a document you own. You can reuse it anywhere: in a Gemini skill, or in the instructions of a ChatGPT or Claude project. Claude has the same idea under the name skills, and our pick of the best Claude Cowork skills shows what they look like.

Frequently Asked Questions

Is Claude's memory available on the free plan?

Yes, since March 2, 2026. Claude's release notes show it reached Team plans in September 2025, then Pro and Max in October 2025, before free users.

Why can't the AI find the file I sent it yesterday?

Because you attached it in a different conversation, so it stayed on that other desk. If you want a file available across conversations, add it to a project's files.