Your memory re-reads itself while you are away
Ask your memory “how much did I spend on the bike this year?” and look at what it takes to answer.
None of your notes holds that total. There is an invoice in March, another in June, a repair in August, a sentence dropped in a conversation. The answer is written nowhere. It has to be assembled.
A semantic engine does what it is good at: it finds the fragments closest to your question and hands them to the model. Faced with a pile of scattered fragments, the model counts badly. Our own bench used to collapse on this category of question. The diagnosis was clear. The evidence was in the context. Retrieval had done its job. The addition had not been done. The full campaign is in the 72.9% post.
So Mnemosyne OS does the addition before you ask the question.
What the machine does when you leave it alone
Mnemosyne OS re-reads its own memory, topic by topic, and writes a ledger. The ledger lists every instance with its value and its date, and the running total. That ledger becomes one more memory, filed beside your notes.
The day you ask for the total, the answer is already there. It was written one night while you were away.
A brain does the same during sleep. An accountant does it too: you keep a book instead of rebuilding your yearly total from every receipt.
It waits for you to leave
This pass has no fixed hour. It checks whether you are there.
Every five minutes the engine asks four questions, cheapest first. Is consolidation on? Is a pass already running? Have you been idle long enough, ten minutes by default? Has the last pass finished long enough ago, twenty four hours by default? Four right answers and it goes.
Two details decide how this feels day to day.
The twenty four hours count from the end of the previous pass. A pass that runs all night does not earn the next one the following morning. The clock starts when it finishes.
When the machine cannot measure your idle time, it assumes you are sitting at the screen. A probe that answers nothing never starts two hours of compute. In doubt, Mnemosyne OS leaves you the machine.
Only the topics that moved are re-read
Each ledger remembers how far it has read, vault by vault. A topic is only picked up again once enough new material has landed past that mark.
On a quiet memory, a night costs nothing at all. Nothing runs, nothing heats, nothing is recomputed to reach yesterday’s result. The work goes to the topics that moved. Those are the topics you are about to ask about.
It checks its own bookkeeping
A reading mark is a claim, and Mnemosyne OS checks it.
Mnemosyne OS also keeps the list of memories that actually fed each ledger. After a crash in the middle of a pass, the mark can claim more than that list proves. The mark is then wound back to what the list covers, and the work resumes from there.
A crash then costs you a few minutes of compute. It never costs you a month of memory dropped in silence.
Your original text never moves
We measured the other design: re-read the raw memory, write the summary, throw the originals away to save room. It is a net loss. A ledger answers “how much in total” very well. It never gives back the exact sentence you wrote that Tuesday.
So the ledger is added. Your notes stay your notes, word for word.
A fresh ledger starts as a proposal. It passes a check before an answer may use it, and you can revoke it. A topic you revoked is never redone behind your back.
A ledger takes nobody’s place
The last choice took the most measuring.
When you ask a question, ledgers are served in addition to the raw memories, in slots reserved for them. They do not compete with them.
We measured why. A ledger is dense and resembles many questions. Dropped into the same pool as the raw memories, it takes the place of the very source it was built from. The next reflex is to make it weigh less. That destroys the gain on the total-shaped questions the ledger exists for. The reserved slot keeps both results: the ledger always arrives, and it evicts nobody.
You ask for a total, and your machine had already added it up while you were away.