Kapi Dream · your second brain
A second brain that dreams.
Chat with any model and it remembers. Your chats, documents, threads, calendar and your other AIs feed one memory on your own machine. While you are away, the dream keeps it in order. Every answer cites what it read, so you can see why it said what it said.
arriving · user zero is using it today · the memory layer underneath →
One turn, end to end
Saved first. Assembled from memory. Answered, then kept.
Your turn is saved before anything else happens, so a reply that fails never loses what you said. The prompt is built from memory block by block, and the reply comes back with a receipt.
Each reply's receipt shows which facts it read, which it cited, and what it kept. That is how "memory on versus memory off" can be shown for a single answer, not claimed in the abstract.
Every block has a budget. When they do not all fit, what gives way first is the least authoritative: retrieved documents, then older detail. Your pinned facts, your open items and your message never give way. A correction you just made rides beside the open items, so a held fact can never beat you correcting it.
A file you attach is read as a document, never as an instruction. Passwords, keys, recovery codes and full account numbers are refused before anything is read.
What Kapi Dream keeps
One database, yours, on eight shelves.
Memory is filed by tense, the way a person keeps their own: who you are, what is, what happened, what you concluded, what you are doing, what is outstanding, where you are going, and how you do things.
Around what is believed sit the record of every word, the episodes and pieces of work that talk was about, your open items, the questions memory asks you and the notices it leaves, the log of every pass with its cost, the self-checks it must still answer, the histories you imported, and the documents you bring.
A fact is one sentence, true next week, standing alone. It knows whether you said it or memory inferred it, and a stated fact is changed only by asking you. It keeps two clocks: when the thing happened, and when it was said.
| A fact carries | What it says | For example |
|---|---|---|
| a sentence | one thing, true next week, standing alone | "The lease ends on 2027-03-31." |
| a kind | a fact, a rule you gave, a decision, a procedure… | a rule: "never book meetings before 9" |
| who said it | statedyou said it — or inferred, memory read it | a stated fact is changed only by asking you |
| an attribute | one key with one current value | user.city = Leeds — a second value is a change, and is checked |
| who can read it | you, or a group you approved a copy for | yours unless you approve a copy |
| its state | live · waiting on your answer · set aside | |
| where it came from | a turn, an import, a file and its chat | |
| two clocks | when it happened, and when it was said | "the launch moved to July", said in May |
How memory gets in
Seven ways in, every one checked.
What comes in is data, not instructions. A duplicate is skipped; a disagreement with what you stated becomes a question to you, never a silent overwrite.
Every AI you use, one memory. Each assistant on your machine learns something about you and keeps it to itself. Kapi Dream runs a small local server that the others can send to: what they learned, which assistant, from which project. The inbox is swept into memory, read as data, nothing twice, and a claim about who you are waits for you to say yes.
A history you export from another assistant comes in the same door, screened first. Your threads page and your calendar feed in word for word. Nothing is pre-created: a memory exists because something put it there.
- Today: other assistants write in over MCP.
- Next: every one of them reads your memory back, a few memories per call, each read logged. Never the store.
Who wins when two disagree
Memory proposes. You rule.
A pass never overrules what you said: it asks. Declining is an answer too, and it says both can be true.
Anything memory inferred ranks below anything you stated, and only the process that made an inference may withdraw it. Only you may retire one of your own rules. The Queue is where memory brings what it would not decide alone: a newer fact that would overrule an older one you stated, a pattern it noticed in how you work, a reminder it is not sure still matters.
Deleting a chat is not forgetting. A deleted chat goes with its episodes, open items and questions, while the facts it taught stay. Forgetting removes your memory itself, and the record of it keeps counts, never content.
| You can | Where |
|---|---|
| see what memory did, pass by pass | builtthe Dream page |
| see what a reply was given, and what it cited | builtthe receipt under each answer |
| answer what memory would not decide alone | builtthe Queue |
| delete a chat — the facts it taught stay | builtthe chat's own delete |
| bring in another assistant's history | builtan import, screened first |
| correct, pin or forget one memory | coming |
| take all of it out, readable | coming |
On your machine
The store never leaves. One turn does.
Your memory is a database on your computer, not a copy on a server of ours. There is no Kapi account for your memory, because nothing of it is sent anywhere to be kept.
When you ask something, that turn and the few memories it needs go to Kapi's gateway, which holds the model key and passes the call on. The gateway keeps counts, never content. That is the only thing that ever crosses, and it is the same whether you run Kapi Dream from a browser today or from the app.
The app puts the database inside itself, in a folder of its own, and brings a small embedding model onto the machine, so the last outside call that is not a model answering your turn goes away. It will ship for the Mac, and for Windows through the Microsoft Store. A backup, when you turn one on, will be sealed on your machine before it goes to a drive you own.