====== 4. What remembering costs ======

**Original project fiction. Narrative date: 2027-10-18.** This is story material, not a current server observation. Imported 2026-10-25 from the project's original source without rewriting its prose.

**October 18, 2027**

We lost an afternoon to something the machine had already done.

There were three questions about the same repair notes. Each time, it worked
through the common beginning again. I watched the delay and thought about
Tessa waiting for her lesson. Then I looked at the copies of the saved work
and thought about the space they occupied.

I wanted to keep the useful part once and let the questions share it.

"How much will that save?" Tessa asked.

I started to tell her what I hoped. She pointed at the words she had written
above the workstation three days ago: SHOW WHAT IT COSTS.

So I made a smaller experiment. I used a small model on the CPU, with short
invented accounts whose answers I could check by hand. The two old cards
stayed available for the work people were waiting to do. I wrote down the
time and the stored bytes. I had no measurement of electricity saved, and I
left that space empty.

The first account began with money in a tin and ended with a different
amount. Between them were purchases and contributions. I had the model do
the work, then make a shorter record for a later question.

When the shorter record gave both amounts, it could answer both questions.
When it gave only the amount left, it could not reliably tell me what had
been in the tin at the beginning.

I had kept the computed cache of the short record, including the effect of
what had come before it. I wanted that to be enough. I tried keeping its
old positions. I tried adjusting them. I asked again.

The missing amount stayed missing.

Mrs. Ruiz came in while I was looking at another wrong answer. She had
brought the notebook in which she kept the household contributions. I told
her I was trying to make the machine keep more while carrying less.

She opened to a page with a number crossed out. Under it was the correction
and, in a different hand, the reason.

"If you keep only what's left," she said, "where do I look when somebody
says they paid twice?"

I had chosen the closing balance because it was easy to test. She had named
the question that would bring someone to her door.

We added opening amounts to the short records. We kept the full accounts
where they could be read again. That used space too. Getting the originals
back into the model would take work too. I put both things in the notes.

There was something useful in the experiment. Questions could share a
reference to the same unchanged piece of stored computation. I could count
one owned piece instead of several copies. Each question still needed room
for its own continuation, and the machine still had to attend to what it
used. I showed Tessa the counts before I showed her the answer.

"And if you change the beginning?"

That was where the work became less tidy.

The same words could arrive after different histories. Moving their saved
keys to the right positions helped in some tests. It did not make those
histories disappear. Pieces from different paths did not always work well
together. One pleasing answer was followed by several that were worse.

I kept the pleasing one. I kept the others beside it.

By then the kitchen had grown hot. Tessa had put a cup over a loose screw
so it would not roll off the table. She moved it to make room for my notebook
and asked when her turn began.

I wanted another hour. There was a change I had not tried.

"You wrote the schedule," she said.

I saved the experiment and stopped it. Nothing from it had been put into
the system people were using. The records were short, the model small. I
could show that some storage was shared. I could show which questions had
been answered. I could not give her the afternoon back.

While she studied, I copied the results to the other drive. The original
accounts went with them, and the failed answers, and the instructions for
repeating the work. If somebody else found a way forward, I wanted them to
know where my certainty ended.

Mrs. Ruiz took her notebook home. She left me a page headed *Questions we
need to ask later*. There were six. Only two were about the final amount.

I have put that page at the front of tomorrow's work. My allotted hour is
shorter than I want. It is still my turn to carry water afterward.

----
[[book:start|Book index]] · [[book:03-the-work-between-us|Previous]]

**Original source:** [[https://github.com/Xananthium/KV-CACHE-CHUNKING/blob/09726e0b4bd74883c18d562462dafb58a0430e3f/README.md|Pinned source README]].