Why you do not have to explain yourself twice, and what is kept where.
Four kinds of thing survive the end of a session, and they are kept apart on purpose.
- How you work. Preferences, decisions and corrections you gave: that you want plain answers, that a certain customer is always addressed formally, that a folder is off limits. Kept in one file and curated rather than piled up.
- What happened. A running log of every notable write, with time, who did it and what it was. This is the record you can check when you want to know when something changed.
- What each specialist learned. Every specialist has its own notes: quirks of a template, an approach that worked, a dead end not worth trying again. Kept per specialist so a design lesson does not muddy a filing decision. A note from work you have not looked at yet is marked as provisional, because the only judge so far was the specialist that did the work. When you later say the result was wrong, the note is struck with the date and the reason instead of quietly staying in force, which is how a wrong lesson would otherwise get more authoritative every time it is followed.
- Each session's hand-off. What was done, what comes next, the reasoning behind it, and any corrections you gave. Written when you close a session, filed by year and month.
Everything is a plain text file you can read and edit yourself. Correct something in there and it holds from the next session on.
The effect builds up. After a few weeks it knows how you want things filed, which wording you keep correcting, which subjects matter to you and which specialists you actually use. That is the difference between an assistant you re-brief every morning and one that has been with you a while.
At the start of a session Zanmai reads your profile and your own contact entry. Before answering anything about past decisions or preferences it goes to memory first rather than trusting what happens to be in the current conversation, because that is exactly where a confident wrong answer comes from.
The specialists' own notes are read when a conversation touches what they cover, not on every start, so the session does not carry ballast it does not need.
Memory that is read at the start of a run costs something every time it is read. Three days of real use produced 678 lines of notes for one specialist, and all of them went into that specialist's context on every job. Left alone that is a file that quietly gets more expensive the more you use Zanmai, and harder to trust: forty rules read at once are not forty rules followed.
So there are two layers, and they are treated differently.
A rule stays. The headline and the bounds, where it holds and where it stops, are what applying it needs, and they are short. What moves out is the paragraph explaining how it was learned, into a dated archive beside the file, with a pointer left behind. When you doubt a rule, the reasoning is one click away; when a job just needs to follow it, it is not in the way.
A rule is never retired by the calendar. A decision like "do not suggest that tool again" has no expiry date, and rotating it out after two months means it gets suggested again in the third. What leaves the live file is what has been struck: an entry your correction disproved is history the moment it is struck, and history belongs where history is kept.
What nobody has checked is asked about, not dropped. A note from work you have not judged yet is provisional. When it is still provisional a month later, Zanmai puts it to you rather than deciding for itself, because the lessons nobody confirmed are exactly the ones that would disappear silently.
The running log is the opposite case. It is only ever appended to and only ever searched, never read whole, so its older months move into an archive beside it with one line saying where they went. Searching still finds them.
Memory is not a transcript. The conversation itself is not what gets kept; what gets kept is what remains true afterwards. A single fact about a subject belongs in that subject's notes, not in memory. A preference about how you want to be worked with belongs in memory.
Closing a session properly is what does the sorting. It writes the hand-off and promotes what should outlive the session into long-term memory, with your agreement rather than automatically.
Claude Code has its own automatic memory, and Zanmai switches it off deliberately, for two reasons.
One fact should exist in one place. Automatic memory would write notes about your material into a folder outside your vault, so the same fact would exist twice with no link between them, and the copy outside would not travel with your notes.
And curated beats automatic for anything meant to last. Automatic memory keeps whatever seems worth keeping mid-conversation. Closing a session is a deliberate moment: what was actually achieved, what follows, what you corrected. That produces a record you can trust rather than a pile you have to sift.
- A working session, start to close, when memory is written
- Contacts, where persistent notes about you live