How AI Browser Extensions Handle Memory Limits
Every AI model has a context window — a hard limit on how much text it can see at once. Here's how well-designed browser extensions stay within that limit and still give useful answers about long or complex pages.
Key takeaways
- Every AI model has a fixed context window — Gemini 1.5 Flash's is 1 million tokens, which covers most pages comfortably, but there are still limits.
- A browser extension has to choose what to send: the full page text, a summary, or a selection — and that choice determines what the AI can answer.
- Browsy sends the extracted article body by default, falling back to chunked extraction for pages that exceed the context window.
- Nothing is cached or stored on Browsy's servers — every request is a direct call from your browser to your AI provider.
On this page
When you ask an AI assistant a question about a webpage, something has to decide what the AI actually sees. The webpage itself can be enormous — a long Wikipedia article, a technical specification, a research paper — and AI models have a hard limit on how much text they can receive at once. That limit is called the context window, and how a browser extension handles it determines whether the AI can answer your question meaningfully or gives a vague, incomplete response.
What the context window actually is
Every AI model processes text in units called tokens — roughly 3–4 characters each in English, or about 0.75 words per token. The context window is the maximum number of tokens the model can hold in memory for a single request. Everything the model needs — your question, the page text, any instructions from the extension, and the model’s own response — has to fit within that limit.
Limits vary widely by model:
| Model | Context window (tokens) |
|---|---|
| Gemini 1.5 Flash | 1,000,000 |
| Gemini 2.0 Flash | 1,048,576 |
| Grok 3 | 131,072 |
| Claude 3.5 Sonnet | 200,000 |
1 million tokens is about 750,000 words — well beyond any ordinary webpage. For most pages, the limit is not a practical concern. The pages where it matters are unusually long ones: a complete transcript, an entire book, a heavily annotated legal document.
What a browser extension does with the page
A browser extension running in Chrome has access to the DOM — the parsed structure of the page including all its text. The challenge is extracting the text that matters (articles, paragraphs, product descriptions) while stripping the noise (navigation menus, cookie banners, footers, advertisements) that would eat into the context window without helping the AI answer your question.
There are three common approaches:
Send everything. Extract all visible text, send it verbatim. Simple, but wastes tokens on navigation menus and cookie banners, and breaks on unusually long pages.
Extract the article body. Identify the main content block — similar to what browsers do in reader mode — and send only that. This is the most efficient approach for articles and documentation.
Chunk and summarize. Split the page into sections, summarize each section, then pass the summaries to the model. Slower, but handles pages of any length.
Most extensions use a combination: article-body extraction first, chunking as a fallback.
How Browsy handles it
Browsy’s content extraction prioritises the readable body of the page — the text a reader would actually care about — using a combination of semantic HTML signals (<article>, <main>, heading hierarchy) and heuristics that filter out navigation, footers, and sidebar content.
For the vast majority of pages, the extracted text fits comfortably within Gemini’s context window. On very long pages — a complete book chapter, a full court filing, an unusually verbose product page — Browsy falls back to chunked extraction, processing sections in order so the AI can still answer questions about any part of the document.
What Browsy does not do: cache the page text on any server, send it to any destination other than your AI provider, or process it in any server-side step that Browsy controls. The content leaves your browser exactly once — as part of the direct API call to Gemini or Grok — and is subject only to that provider’s privacy policy.
When context windows matter in practice
For most questions — “summarise this article,” “what is this product’s return policy,” “what does this term mean” — the context window is not a concern. Where it can affect answer quality:
Long comparisons. Asking “compare section 3 and section 7 of this document” requires both sections to be in context simultaneously. If the document is very long and the extension only sends a portion, the AI may have access to one section but not the other.
Very long transcripts. A 2-hour video transcript can run to 30,000 words — well within Gemini’s limit, but enough to constrain older or smaller models.
Pages with substantial boilerplate. A legal site might include 10,000 words of standard terms and navigation before the clause you actually want. An extension that sends the raw page text will waste half the context window on boilerplate.
A question worth asking about any AI extension
Before trusting an AI extension with pages you care about, check two things: what text it sends and where that text goes.
Open your browser’s DevTools (Network tab) while using the extension and make an AI request. You should see a single request going to the AI provider’s domain — nothing else. If a request goes to an unknown intermediate domain before the AI provider, the extension has a server layer that sees your page content, regardless of what its privacy policy claims.
The architecture is the privacy guarantee, not the policy.