Chat, voice, memory and scheduled runs

What the assistant does in Chat

Ask about the records and it reads them before answering rather than relying on what was said earlier, because the data may have changed. It can add and change records and prefers the app's named actions when one fits, since those carry the app's own rules. Deleting is irreversible, so it confirms before it deletes. Every tool it uses is shown in the conversation as it happens, so a change to your data always leaves a visible trail.

It answers briefly, in a panel beside the app rather than a chat window, and does not narrate its own steps. If something is not in the records it says so in a sentence. It cannot look up a person by a rough name: matching is exact, so for "do I have butter" it reads the whole list and looks, which for a personal app is quicker than guessing spellings.

Voice is the same conversation, spoken

A live Voice call is a way into the same Chat conversation, not a separate one. You speak, the assistant answers aloud in short sentences, and you can interrupt it mid-answer. The records it reads, the actions it can run, and its memory are exactly those of typed Chat, and what you say in a call is remembered the same way, so a preference mentioned aloud comes back in text.

A call needs the microphone, which the browser asks you for; it needs allowance left on your plan; and one call runs at a time. A spoken confirmation follows the same rules as a typed one: an action that needs approval still parks as a proposal card, and saying yes does not approve it. An app can choose the voice it speaks in, or leave that to the voice you picked in Kaddy's settings.

Asking for a change to the app

Tell Chat what should be different and it hands a brief to Code, the build conversation, then tells you in one line that it has done so. It does not describe files or ask you to edit anything. The build runs on its own, reads the Chat conversation itself for context, and lands its result in the working copy, where the App view previews it. Nothing is live until you press Apply.

For something broad, Chat asks first. One question at a time, about the problem rather than the features, until it can summarise the smallest useful version and ask whether to build it. That interview happens in Voice too, in short questions that are easy to answer aloud. A precise request that already says to go ahead skips the ceremony.

Chat knows what Kaddy can and cannot build, so it will say when something is out of reach rather than send it to Code to discover. The capabilities page is the same list.

What the assistant remembers, and how to correct it

Each app's assistant keeps a memory of what it has learned from Chat: that you cook for two, that short books suit the commute. It is a set of readable notes, not a transcript, and a small standing part of it is in front of the assistant on every turn while the rest is searched when needed. It belongs to the app, so starting a new Chat conversation does not start the memory over, and it does not travel to other apps.

In the background the assistant tidies what it has learned: recent conversation becomes dated observations, and observations that recur become notes. It may propose a change to its own standing instructions, but that is only ever a proposal, shown in the Memory view with Approve and Dismiss, because an automatic edit to its own rules would compound into every later answer.

You can read and edit all of it in the Memory view. When an answer keeps repeating an assumption that has stopped being true, correct the note there and say how long the new situation lasts. Memory is not versioned with the app: returning to an earlier version changes the screen, not what the assistant knows. The memory guide works through an example.

Screenshots

You can paste screenshots into a conversation in the common image formats; Kaddy shrinks them before sending. Only a model that can see images receives them. If the app's chosen model cannot, the picture stays in the conversation and a note says which models can, rather than the assistant describing a picture it never saw. Other attachments, including PDFs, are refused rather than converted.

Scheduled runs

An app can take one turn on a schedule: every Sunday at nine, say, with a standing message such as "plan the coming week's dinners". At that time the assistant runs the turn as if you had typed the message, and the reply lands in Chat like any other. That is the whole delivery: you open the app on Sunday and the thinking is done. There is no push notification or email, on purpose, since letting an app think on a timer is a different decision from letting it send mail.

The schedule is set in the app's Settings with a time zone, which matters more than it looks: a Sunday morning in the wrong zone is a Saturday night. An action the scheduled turn requests follows the same approval rules as any other.

Choosing a model, and keeping several conversations

Each app chooses its model in Settings, one for Chat and one for Code, from a menu that says where each model runs: within Kaddy's own hosting, or with an external provider that then receives what you send. Nothing is switched to an external provider silently. The privacy policy describes that boundary.

Each kind of conversation can hold as many sessions as you like. A session keeps its own context and is named from the first thing said in it; rename it to keep it findable, clear it to empty it, or delete it. Memory is the app's, so a new session does not forget anything.