How a saved post becomes an answer.

It starts with you tapping the little bookmark on an Instagram post. That post drops into your saved list, which is a box only you can open.

Instagram is happy to hold that box. It is not happy to hand you the contents. There is no button that says "give me everything I ever saved," and there is no promise that a post you saved will still be there next year.

So the whole rest of this page is about getting your own stuff out of that box and into somewhere that answers to you.

A small Chrome extension sits in your browser. Every four to seven hours it wakes up, opens an Instagram tab, and asks Instagram for one page of your saved list.

Here is the trick. The extension does not ask from outside. It runs its question from inside your own Instagram tab, so Instagram sees a normal logged-in you asking a normal question. No password is copied anywhere. Your login cookie never leaves your browser.

Instagram answers with a page of posts and a bookmark saying "next page starts here." The robot follows that bookmark until it runs out of pages.

Because being clever is how robots get things wrong. The extension does not keep a list of what it already sent. Every single run, it walks your whole saved list and sends every link again.

That sounds wasteful, and it is a little. But it means the extension can never be confused about what it missed, never get stuck, and never quietly skip something. If it gets sent twice, the other side just shrugs.

The database is the one keeping track. It has a rule that says "one row per link," so the second, third and hundredth copy of the same link change nothing.

Every batch the extension sends carries a secret token, like a wristband at a concert. The server checks the wristband before it even looks at what is in the box.

That order matters more than it sounds. If the server read the box first and checked the wristband second, a stranger could get their junk halfway in before being stopped. Checking first means a wrong token can never create anything at all.

Once you are through the door, each link is looked at on its own. A broken one gets marked broken and the rest of the batch keeps going.

A link is not a memory. Links rot. So the moment a post arrives, ChronoCast goes and fetches the actual thing: the caption, who made it, how long the video is, the video file itself, and the thumbnail. All of it lands in private storage that only you can open.

Then it listens to the video. A speech-to-text model runs on the computer, not in the cloud, and writes down every word that was said out loud. That transcript turns out to be the most valuable part of the whole capture, because most of what a video says is never in the caption.

Each of those steps is allowed to fail on its own. If the video is private and cannot be downloaded, the caption still gets saved, still gets a transcript if one exists, and still becomes findable. A half-remembered thing beats a blank.

This is the part that makes the whole thing work, and it is genuinely strange. A small model reads the title, the caption and the transcript, and turns all of it into a list of 1,536 numbers.

Think of those numbers as an address in an enormous invisible city. Everything about books ends up on one street. Everything about football ends up across town. Nobody told it where the streets are. It worked them out by reading a very large amount of text.

The magic is that your question gets an address too. So finding things that mean the same as your question turns into finding neighbours, and finding neighbours is just arithmetic. Fast, boring, reliable arithmetic.

After the numbers are done, a small fast AI reads the capture once and leaves three sticky notes on it: a one-sentence summary, up to eight topic tags, and a list of the people, tools and places it mentions.

Those tags are what you click on to filter your library, and the names are what get wired together into a map of everything you saved.

One honest catch, and it is a real one: the tags are written after the 1,536 numbers are already made, so the tags are not part of the address. A post tagged "book recommendations" is not findable by meaning because of that tag. It is findable because of what the video actually said.

Two searchers go looking, at the same time, and they are good at opposite things.

The first one searches by meaning. It turns your question into an address and grabs the nearest neighbours. It will happily find a video about novels when you asked about books, because it does not care about the letters.

The second one searches by letters. It looks for your exact words sitting inside the captions and transcripts. It is useless at synonyms and unbeatable at names, handles and weird spellings.

Then their two lists get merged. Anything both searchers liked floats to the top, because two independent yeses is much stronger evidence than one loud one.

There is a line drawn at 0.30. If the closest thing in your whole brain is further away than that line, nothing goes to the AI at all. You get told, plainly, that there is nothing in there about that.

This is the most important rule on the page and it is the one people skip. An AI handed weak, half-related scraps does not say "these are weak." It writes a confident, fluent, wrong answer, and it puts footnotes on it.

Where should the line go? We measured it instead of guessing. Sixty-four real questions were written down, along with which saved posts should answer each one — including fifteen this archive genuinely cannot answer, about beekeeping and toothaches and Japanese grammar. The fifteen it cannot answer never scored above 0.284. The ones it can never scored below 0.413. So the line goes in the gap between them, and on that test it opens for every answerable question and refuses all fifteen of the others.

Only what survived. Up to eight captures, each one numbered, each one carrying its title, its author, your note and up to four thousand characters of its transcript. That is it. No web search, no memory of your other chats, no general knowledge allowed in through the side door.

The numbering is the part to notice. The list the AI reads and the list of sources you see are built in one pass with one counter, so footnote [1] cannot ever point at the wrong video. If they were built separately they would agree by luck, and luck runs out.

Then it writes the answer, and every claim carries the number of the capture it came from. Click the number, watch the actual post it read. That is the entire promise: not a smarter answer, a checkable one.