Without memory
"Hello. How can I help?"
You pause, then retell the whole three weeks. Halfway through you run out of steam. It is not that it is not smart; it starts from zero every time.
Scenario / Companion
Three weeks of talking,
and then it forgets
You talked with it for three weeks. You mentioned your cat, said you had not been sleeping well, and asked it to hold on to one small thing for Friday. Then five days pass, you come back — what is the first thing it says?
01
That night
Same you, same subject; the only difference is whether it remembers.
With recollect
"Last week you said the project was stuck and you wanted to stop for two days. Are you feeling better today? And how did the cat thing turn out?"
It does not replay everything you said. It picks the few things that are useful right now. That is what not forgetting looks like: quiet, but it picks up the thread.
02
What it keeps is not a chat log
It keeps four things it can use directly. And they are not only in its own encrypted notebook (core).
01 / Who you are
Profile and how to address you
How to call you, what you care about, which topics to leave alone. This half lives in its own core, visible only to it.
02 / How you felt then
Mood and state
"Exhausted" is a record with a timestamp and an expiry, not a permanent label stuck on you.
03 / How it turned out
Events and causality
Which day it happened, why, and how it ended. It can go back to that turn, not just say it thinks it happened.
04 / What you asked it to keep
Notebook
A passing "remind me on Friday" is append-only; the summary feeds into the next day's opening.
Do not read this as memory living only in core. Core is its own half. What the two of you go through — time, events, causality, how the mood shifted — is recorded in world memory as well, ready to be pulled up when needed, instead of staying in the context of a single session.
03
One overnight: from "I'm tired" to tomorrow's first line
Click the four moments to follow one sentence. The case is illustrative; the mechanics follow the real implementation.
04
What it does not do
Smaller claims are easier to trust.
It does not invent preferences
No evidence means the field stays empty; it does not ask whether you like something and then fill in an answer itself.
It does not turn mood into fact
"Tired" is how you felt then, with an expiry; it never becomes "health: poor".
It does not pass inference off as your words
"because the Raspberry Pi would not run it" is an inference, marked as such, never passed off as something you said.
It does not share with anyone else
Persona, profile and notebook belong to this agent alone; host and account details are isolated too.