ReMyndwhat is a context window?
a context window is everything a model can see while it writes one answer, measured in tokens.
anthropic's documentation puts it plainly: the context window "refers to all the text a language model can reference when generating a response, including the response itself", and it works as "a working memory for the model". the important word is working. it holds for one request and then it is gone.
as of september 2026 the top end is 1,000,000 tokens, which is large enough that people have started treating it as memory. it is not memory, and that mistake is behind most of the frustration people report with ai tools.
what counts against the context window
everything in the request counts, and the list is longer than most people expect.
- the system prompt, including whatever the product prepends without telling you.
- every message in the conversation, both sides, all the way back.
- tool definitions, which on an agentic setup can be thousands of tokens before anything happens.
- tool results, including the full text of files an agent read.
- images and documents you attached.
- the model's own reasoning, where extended thinking is on.
- the answer being written, because output shares the same budget.
anthropic's docs are explicit that cached content still occupies the window: prompt caching "changes what you pay for those tokens, not whether they count".
for scale, openai's embedding models cap a single input at 8,192 tokens, which is roughly a long blog post. a context window is bigger than that by two orders of magnitude and still smaller than a week of your work.
how big are context windows in 2026
the headline numbers cluster at two sizes, and the gap between them is 5x. windows were 4,096 tokens in 2022 and 200,000 by 2024, so the direction is clear even if the ceiling is not.
| model family | stated window | source |
|---|---|---|
| claude, current top end | up to 1,000,000 tokens | anthropic docs |
| claude sonnet 4.5 | 200,000 tokens | anthropic docs |
| gemini, many models | 1,000,000 or more tokens | google docs |
the token limit is a hard edge, not a soft one. send more than the window holds and the request fails outright rather than quietly dropping the middle.
a million tokens sounds infinite until you convert it. google's long context page offers the conversion: 1,000,000 tokens is roughly "50,000 lines of code (with the standard 80 characters per line)", or "8 average length English novels", or "transcripts of over 200 average length podcast episodes".
that is a lot of a book and almost none of a working life. a single week of one person's screen, transcribed, is larger than 8 novels.
why bigger windows still forget
the failure mode is not truncation. it is degradation, and it starts well before the limit.
anthropic's context window documentation names it directly: "as token count grows, accuracy and recall degrade, a phenomenon known as context rot. this makes curating what's in context just as important as how much space is available."
the effect was measured before it was named. the lost in the middle paper, published in july 2023 by nelson liu and colleagues, found that performance "is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models".
so a 1,000,000 token window is not a 1,000,000 token memory. it is a large surface on which position still matters.
context window versus memory feature
these are different mechanisms and the products blur them on purpose.
| context window | memory feature | |
|---|---|---|
| lives | inside one api request | in a store outside the model |
| survives a new chat | no | yes, if the product retrieves it |
| size | 200,000 to 1,000,000 tokens | unbounded, but only a slice is used |
| who fills it | the product, on every call | the product, from prior conversations |
| covers your work | only what you pasted | only what you said in that product |
the last row is the one that matters. chatgpt, claude and gemini can each remember what you told them. none of them can remember what you did, because none of them saw it. we take that apart in ai memory compared.
what actually fixes the limit
the answer is not a bigger window, it is a better index.
- keep the corpus outside the model. documents, history, notes, whatever it is, stays in storage.
- search it per question. find the handful of passages that are actually relevant.
- put only those in the window. a few thousand tokens of the right thing beats 500,000 tokens of nearly right.
- cite what you retrieved, so the answer is checkable rather than plausible.
that pattern is retrieval augmented generation, and it is why anthropic's engineering write up on effective context engineering treats context as a budget to spend rather than a bucket to fill.
it also explains a practical result people find surprising. a smaller model with a good index usually beats a larger model with none, because the retrieval step is doing most of the work.
the corpus nobody has
there is a hole in step one, and it is the reason this article sits on this site.
retrieval only works over something you can search. your files are searchable. your email is searchable.
the dashboard you read on tuesday, the slide someone shared on a call, the terminal output you scrolled past, the ticket comment you skimmed: none of that is in any index, anywhere.
so when you ask a model about your own week, it has nothing to retrieve. the window is not the constraint. the corpus is missing entirely, which is the argument in context engineering and the reason spotlight search cannot help either.
where remynd sits
remynd exists to build that corpus on a mac. it records the focused window, runs optical character recognition on device, and keeps the resulting text in a local index, so there is something to retrieve when you ask.
the retrieval shape matters here. remynd does not pour your history into a context window. it searches the local index, pulls the relevant slices, and sends only those to the model, which is the same discipline the anthropic guidance describes.
the scope, stated exactly: capture is the focused window, ocr and storage stay on your mac, recordings default to 30 days of retention, and you can exclude specific apps or sites from capture entirely. asking a question sends the retrieved slices to a cloud model, and there is no local model option today. optional backup is encrypted and exports to your own s3 bucket. the honest line is that the archive lives with you, not that nothing ever leaves your machine.
for the retrieval mechanics, read what is rag. for what it looks like when an agent uses it, read how to give an ai agent context.
download remynd free for mac → free to download, runs locally, no card.