Docs/Conversations/Chatting with the AI

Chatting with the AI

Conversations let you think out loud with an assistant that knows your notes. The panel has your open conversations on the left and the active chat on the right. Instead of quietly reading everything you've ever written, the assistant searches your notes, follows links, or looks something up on the web only when it actually needs to — and tells you what it found.

Opening a conversation

The conversations panel sits below the editor and stays out of sight until you call for it. There are a few ways to start:

Show or hide the panel

View → Toggle Conversations (⌘⇧K) opens or tucks away the panel. Hiding it leaves everything in place — your conversations are waiting when you bring it back.

Start fresh

The New Conversation button above the editor, the + button in the conversation list, or Learning → Ask a Question. This opens a blank conversation with no particular note attached.

Ask about a note

Right-click a note's tab above the editor and choose Ask About This…. This starts a conversation focused on that note, which the assistant knows is your starting point.

Close a conversation

The × on a conversation puts it away for good, so you'll be asked to confirm first — a closed conversation can't be reopened.

"This note" follows you

A conversation started from a note isn't locked to it. If you switch to a different note while chatting, the assistant keeps track of both — the note you started from and the one you're looking at now — so when you say "this note," it knows which one you mean.

How the assistant knows your notes

The assistant doesn't read your whole knowledge base every time you send a message. Instead it reaches for what it needs, when it needs it, and can:

Search your notes

Look across everything you've written to answer "what have I said about this?"

Read a note

Open a specific note in full when it needs the details.

Explore connections

Answer questions about how your notes relate — what links to what, which notes share a tag, which claims cite a given source.

Search the web

Look up facts, current events, or documentation that isn't in your notes, and read a specific page in full.

The only thing the assistant always knows up front is today's date and which note you're working with. Everything else it goes and finds on purpose.

Reading is free; changing is not

The assistant can read and reason across your notes freely, since reading changes nothing. But it can never edit your knowledge base on its own. When it wants to add or change something, it offers a draft you can accept or discard — see The propose → review → approve loop.

Sending a message

Type in the composer and press ⏎ to send, or ⇧⏎ for a new line. Starting a line with / brings up a menu of quick commands and thinking tools you can run in the chat — for example, a command to tidy up a long conversation into a shorter summary. If voice input is turned on, a mic button next to Send lets you dictate your message out loud. Your message appears in the transcript right away, so you can see it while the assistant works on a reply.

Watching a reply come in

Replies appear a little at a time as the assistant writes, with a small "thinking" indicator that stays put during any pause — including while it's off searching your notes or the web. Formatting shows up as it goes, so nothing jumps around when the reply finishes. While a reply is in progress, the Send button turns into Cancel so you can stop it at any time.

Choosing a model for one conversation

Above each conversation is a model picker. It starts on Default — which names the model your AI settings currently point at, so you can see what "default" actually means without going to look — and switching it affects only that conversation. Two conversations open side by side can be running on different models from different companies.

The list is grouped by who makes each model: Claude, GPT and the o-series, Gemini, and any local models you've added. You'll only get an answer from a service you've set a key for; pick one you haven't and Minerva says so and offers to open Settings rather than failing quietly.

Beside it sits an effort picker, when the chosen model has something to offer there:

It only appears when it means something

Models with no thinking control of their own — the quick, cheap ones — simply don't show the picker rather than showing one that does nothing.

The levels follow the model

Each model offers what it actually supports. Switch a conversation from a model with five levels to one with three and a setting the new model can't honour drops to the nearest one it can, rather than being silently ignored or erroring at send time.

Default names its inheritance

Leaving effort on Default shows the level it's inheriting from your settings, so an unset conversation is still a known quantity.

Switching mid-conversation is fine and often useful: think something through with a fast model, then hand the hard part to a more capable one with the whole conversation already in front of it.

Citations and references

When the assistant uses the web, each source it relied on shows up as a numbered note beneath the reply, with the title, the site, and the exact passage it quoted. Click one to open the page in your browser. A cite action next to it saves the source to your library and drops a citation into whichever note the conversation is tied to — the same way you'd cite a source anywhere else in Minerva.

When the assistant refers to your own notes, it simply names them in its reply, so you can jump straight to what it was reading.