Skip to main content

RAG attention, the slider that decides who answers

Every vault carries a "RAG attention" slider, from 0 to 100 percent. It sets how much that vault weighs when memory is queried: at 100 it counts fully, lower down its results weigh less against the other vaults, and at 0 it stops answering entirely.

The weight says who answers, never who records​

A vault at 0 is mute in answers: it is not queried, and none of its content reaches the prompt. But it keeps recording normally. Its watched folders stay watched, a file that changes is still ingested, and the day you raise its slider it answers with everything that arrived meanwhile.

Reading and writing are two separate acts, and the slider only ever meant the first.

Solo mode​

From the far view of the Vaults window, picking a vault sets that one to 100 and every other one to 0. An "All" button brings everyone back to 100.

Both gestures overwrite a mix you set by hand: if you had tuned 70 / 30 / 100, it is gone. That is the price of a one-click switch.

When the weights form a mix that neither button describes, the display says "mixed" rather than pretending everything is on. When everything is at 0, it says "all muted".

The chat's scope says the same thing​

The "Scope" pill at the top of the chat and the attention slider set the same fact: which vaults answer. They are not two settings, they are one, seen from two places.

So soloing a vault in the Vaults window shows up in the chat immediately, which names it. And the other way round: ticking vaults in the chat's pill puts them back to 100 and the others to 0, with the Vaults window and the neural map moving along. The most recent gesture wins, wherever you made it.

"All" wakes every muted vault, but a mix you had tuned by hand between the vaults you keep is preserved.

Two situations cannot be written as a list of vaults, and the pill names them instead of saying "All": "All but 3" when a few vaults are muted among many, and "No vault" when everything is off — in which case no answer will have any memory until you turn something back on.

Reopening an old conversation does not restore the scope it was saved with: reading a conversation is not a choice, and it would silence vaults you may have just turned back on. Only "Off" is restored, because nothing else could ever bring it back.

Letting Mnemosyne OS pick the vaults (Auto)​

The chat's "Scope" pill has an "Auto" position. It means: for each question, the model looks at the list of vaults and picks the ones the question is about, instead of staying on your last setting. You allow it by choosing that position; nothing happens on its own.

Three kinds of vault are never offered to it: those under MAXIMUM protection — a journal is read when you name it, never because a model thought it fitted —, those whose attention is at 0, which you silenced and this mode has no business undoing, and the vaults belonging to apps. To read one of them, tick it yourself in the list.

When the model cannot decide, or when that call fails, every vault is read, as usual. So this mode cannot cost you memory: at worst it aims at no less of it.

Under the answer, the badge carries a spark when the vaults were picked by the model, and says which ones; if it named none, it says every vault was read. You can always tell who chose.

What it costs: one extra question asked of the model before each answer, so a little waiting and, on a cloud route, a few tokens.

What the slider does not do​

It protects nothing. A vault at 0 is not private, it is silent: name it explicitly in a question and it answers. To keep content from reaching a cloud provider, it is the vault's protection level you want, not its attention.