The index first, then the right page.

1. The AI picks the words. From your question it takes two or three keywords, not the whole sentence. «When did we change the courier for Rossi?» becomes courier Rossi.
2. The brain looks in its index. It does not reread the notes one by one: it has a ready-made index of every word, like the index at the back of a book, updated every time someone writes. It searches file names too. If all the words together find little, it tries again with «at least one of these».
3. A list comes back, not the notes. For every note found: where it is, when it is from, and the two or three lines where the words appear. Like search-engine results: a title and two lines under it. Notes matching only some of the words sit at the bottom, under a line that says «maybe».
4. The AI opens the right note, all of it. From the list it picks the one that answers, looking at the date too, and reads it from the first line to the last. That is where it finds the answer, not in the excerpts.
5. If needed, it takes the whole file. When one note is not enough, it asks for the whole folder: that client's rules, its card, the latest decisions and the most recent notes, in one block. And the block names what did not fit, so the AI can ask for it.

«When did we change courier?» is an index question: search, open the note, done. «We are working on Rossi» is a file question: before doing anything the AI needs that client's rules, what was decided, what happened in the last few weeks. No keyword search gives it all of that at once.

That is why every client, project or person you follow is a folder with a card. When the AI opens the folder it gets the rules first, then the card, then the decisions, then the notes from the most recent. What weighs too much stays out but gets named: an eighty-page note is opened on its own, when needed, not on every question. What goes in, and in what order, is on the context costs page.

One note per working session, with the day in the name and a title that states the fact: «Courier changed: express, never on a Friday», not «Update». It is what you would do in a paper archive, and for the AI it is gold: it reads the date from the name, the fact from the title, and the rest from the text.

There is no AI rereading everything on every question to «understand» what is needed: the index updates as you write and a search is an index lookup, fast and free. It is also why the memory stays yours: they are Markdown files in real folders, you open them with any program and take them away whole.

Searching is search, opening a note is read_file, taking the file is load_context, and the first look at a client is start_session. They are four of the seventeen documented tools, and your AI picks them on its own.