You have an AI assistant that actually knows what you own, where you are, and what you've talked about before. It's like having a personal concierge who never forgets and never judges your snack habits.

#What It Does

The Chat tab is your direct line to a context-aware AI assistant. Unlike a generic chatbot that starts every conversation from scratch, your ambientChat assistant has access to your full inventory, your documents, your location context, and your conversation history. Ask it about your stuff, your receipts, what's expiring, or what you talked about last week — it already has the background.

It can answer questions, generate shopping lists, summarize receipts, estimate the value of your collections, and help you make sense of everything you've cataloged. Think of it as the one friend who actually remembers where you put things.

#How to Use It

#iOS App

  1. Tap the Chat tab at the bottom of the screen
  2. Type your question or request in the text field at the bottom
  3. Tap the send button (the ↑ up-arrow) to submit your message. On iOS, the keyboard's Return key adds a new line inside your message instead of sending it — so you can write multi-paragraph or list-style prompts and only send when you tap ↑. While the AI is answering, the ↑ button turns into ■ Stop. Tap it at any point — even before the first word appears — to end the answer: whatever had arrived stays, and an answer with nothing yet simply goes away. If you stop before any AI was asked, the chat says Stopped — not sent under your question, with a Send again button — nothing ran and nothing was saved, so there is nothing to look for in Q&A.
  4. Your AI responds in the conversation area above, with your messages on the right and AI responses on the left

The Context Bar: At the top of the chat, you'll see a bar showing the current context being sent to the AI — your active beacon or location, nearby items, and any focused context. This is how the AI knows where you are and what's around you.

Attachments: Tap the attachment button (paperclip icon) next to the text field to send photos and images, PDFs, Word/Excel/PowerPoint files (.docx/.xlsx/.pptx), notes (.md), plain text, and spreadsheets (.csv/.tsv) along with your message. You can also pick something you already own — Pick from myAI Context attaches an existing inventory item or document by reference.

Pick from myAI Context reaches the whole folder — iOS and Web. The picker used to load the first 200 documents (or items) in a folder and stop there without saying so, so in a folder with more than that, the one you were looking for might simply not be listed. It now loads a screen at a time as you scroll, all the way to the end, newest first, at every level you drill into.

Attach a contract, a deck or a spreadsheet and just ask. Word, Excel and PowerPoint files are read for their text — the app extracts it and hands it to the AI, so you can ask "what's the notice period?" about a .docx the same way you would about a PDF. Choose Save as document and the original file is kept in Documents too, searchable straight away; choose Just discuss and it's used for that conversation only.

Two things worth knowing: old .doc/.xls/.ppt files need saving as .docx/.xlsx/.pptx first (see Documents), and a file that can't be read — password-protected, corrupt, or a slide deck that's all images with no text — will say so rather than letting the AI answer as though it had read it. If you attach something chat can't take at all, it tells you which file and what to do instead.

Choose what happens to an attachment. Once a file is attached, pick one of three:

Choice What you get
Just discuss Nothing is saved. The AI can see the file for this conversation only. For a spreadsheet (CSV or Excel) that means the whole sheet, not just its first rows — ask about any row, or ask it to add the rows to your inventory, and the spreadsheet itself still stays out of Documents unless you say "save it". A photo, PDF, Word or PowerPoint file or note is kept the same way, so later in the conversation you can still ask "what was in that photo?" or "what did that contract say about notice?". They last until you trash the conversation, or 30 days.
Save as document The file is filed in Documents, in a folder you choose. A notice says so; attaching the same file again uses that copy instead of saving another. Later in the same conversation the AI still knows the file is there — "now import it" or "what did page 3 say?" works without attaching it again.
Add to inventory The file is scanned and any products in it become inventory items, filed in an inventory folder you choose. Your photo is kept alongside them, and later in the conversation the AI still knows what you attached and what the scan added from it. For a spreadsheet, the AI shows you which column it will use for the name, brand and so on, then adds every row as an item — the spreadsheet file itself is not saved.

It's one choice, not a set of checkboxes — documents and inventory are separate folder trees, so the folder picker below shows only the folders that can hold what you're saving. Pick "Add to inventory" and you browse inventory folders; pick "Save as document" and you browse document folders. Switching between them resets the folder, because a document folder isn't a valid home for an inventory item. You can drill into subfolders, and New Folder creates one of the right type in the spot you're looking at.

Attach a photo and ask the AI to save it. Send a picture with something like "add this to my inventory" and the item is created from what the AI can see plus anything you tell it — and your photo is kept as the item's image, rather than the AI illustration scanned items normally get. Handy for one-of-a-kind things where the actual picture is the point. The same works for something you already own — "this is a better photo of my cordless drill" attaches the picture to that item.

Answers hand you the actual thing. When the AI touches something you own — creates an item, updates one, finds a few, opens a document — the answer carries the real record, not just a sentence about it. What you see scales with how many there are:

How many What appears
One The item's actual inventory card — the same card you'd see in myAI Context, with its photo, name, and details
Two to six A row of thumbnails you can tap
Seven or more A single chip — "24 items" — that expands in place to list them, grouped by kind, so you stay in the conversation

Back from an item takes you to Inventory, not back to chat. On iPhone, tapping an item's card or thumbnail opens it in myAI Context with Inventory beneath it, so < lands on Inventory with its category bar, ready to keep browsing. To return to the conversation, tap the Chat tab; the thread is still there. And wherever you are in myAI Context, tapping its tab takes you straight back to its start screen, with the folder you had open still open.

Back from a document takes you to your Documents list. Tap a document's card or link in an answer — on iPhone or on the web — and < goes up to the Documents list rather than back to the chat. (A picture the AI just made for you still goes back to its card in the chat.) On the web, your browser's own Back button still returns to the conversation. A document you open from your Q&A list goes back to Q&A, where you were.

One card tells you where it's filed. When an answer cites a single item or document, its card names the folder it lives in — Taxes, say, or Home if it isn't in a folder — and reads it fresh, so a document you've just moved shows where it is now.

On Desktop AI, the cards arrive as the answer finishes. An answer from your Mac links each record it talks about by name in the text, and those records appear as cards under it once the last words are in, without a heading over them. If an answer forgets the links, the things it names still get their cards. Desktop AI answers are also kept short: each thing is named once with what you asked about, such as when it expires, because the card under the answer opens the rest with a tap, and your Mac writes more slowly than a cloud model. A cloud model can show its cards while it's still answering, sometimes in titled groups. The cards are the same either way, and so is reopening the thread later. Desktop AI works this way because your Mac answers sooner when it doesn't stop partway through just to show cards.

Anything an answer added or changed always gets a card. If the AI created an item, saved a document's text or changed a record along the way, that record's card is under the answer even when the answer itself doesn't mention it, on every model, on the phone and the web, in Q&A, and when you reopen the conversation. So nothing the AI did can slip into your library unseen. The one exception for now: a change to a list someone else shared with you (ticking a line on their shopping list) gets no card, because the card would open your own library, where their list isn't. Ask "what changed on the list?" to see it, and it can still be undone.

Tap any of them to open the real record. If something has since been trashed it shows up muted with an In Trash badge instead of a dead link (a folder the AI deleted for you shows the same way, and a tap takes you to Trash to bring it back), and if it was deleted for good you'll see its name as plain text with a note. The names and pictures are looked up fresh every time, so a thread you reopen next year shows what things are called now, not what they were called then.

This works for photos too. Attach a picture, choose Add to inventory, and the card appears in that same answer as soon as the scan finishes — usually a few seconds after the AI stops typing. You don't need to reload or go looking for it.

Ask AI about a specific item or document: You don't have to start in chat. Open any of your items (the speech-bubble button on the item's photo) or any of your documents (the speech-bubble icon in the header) and tap Ask AI — a fresh conversation opens with that item or document already attached. Type your own question ("Is this contract missing anything?", "Is this still safe to eat?") and send. Because the attachment is part of the conversation, the AI keeps it in scope for follow-up questions, and you can switch models mid-thread for a second opinion on the same attachment.

Voice Input: Tap the microphone button — in the row of round buttons above the message box, between Image Studio and the model button — to speak your question instead of typing. The app transcribes your speech and sends it as a message. Handy when your hands are full of groceries you're about to catalog.

Voice Output: The AI can read its responses aloud. When a response arrives, you'll see play/pause controls to start and stop audio playback. Useful when you're cooking and need hands-free answers about that expiring cream cheese.

Model Selection: Tap the model button at the right-hand end of that same row to open the model selector. The button shows the provider's logo and the model's price band — $ Everyday, $$ Thoughtful, $$$ Premium, $$$$ Frontier — and the selector spells out the band names beside the same $ signs. (Long-press the button to see the band name without opening the selector.) You can switch between AI providers — OpenAI, Google Gemini, and Anthropic Claude — depending on which model you prefer or which handles your question best.

Usage Indicator: A small indicator shows how many tokens you've used against your subscription tier's limit. Keep an eye on this if you're on the free tier — your AI is helpful, but it does have a budget.

Watching it work. While an answer is being prepared, a line under it says what is happening, changing in place. First Finding relevant contextual information…, then Gathering your context: 1,387 items, 3 documents, Study — your inventory, the documents it is reading in full, the zone you're in. A lookup says what it is looking for (Searching the web for Tony Blair's childhood…, Searching your inventory for batteries…, Reading theverge.com…), and Going through what it found… while the AI reads the result. A cloud model that is still working out its answer says Thinking it through…. On Desktop AI your Mac shows Your Mac is reading your context with a thin bar that fills as it reads — Your Mac is catching up on this conversation on a follow-up, Your Desktop AI is reading what it found after a lookup — and on a long read, roughly how long is left (— about 30 seconds). Then Writing the answer… until the first word. When the AI says what it is about to look up — Let me check what's expiring soonest… — that note shows on this line while it works, not in the answer: the answer you read, hear, find in Q&A and export is the answer alone. (If the AI tells you something first and only then says it will check — Your list has 3 items. Let me add the milk. — that first sentence is part of the answer and stays.) A quick answer appears in the bubble from its first sentence. (Answers saved before November 2026 may still start with such notes.) In the coach bubble the note shows in the same way until the answer begins. A step only appears if it lasts more than a moment, so a quick answer goes straight from the dots to the words — nothing is slowed down to show it.

What an answer used. Under every finished answer a small grey line — its receipt — says what of your context it used, where it ran and how long it took: Used 1,387 items · 3 documents · Study · on your Mac · 9.8s (a cloud model shows what it cost instead of on your Mac). Tap it to see the documents it read — tap one to open it — and the technical numbers: 25312 tokens · first words 1.5s · done in 9.8s. The receipt is saved with the answer, so it is still there when you reopen the conversation, on your phone or the web, and in Q&A — where it works the same way, tap and all. An answer that appears the instant you tap AI on a Radar or Lookup card was prepared while you looked, and carries its receipt too — what it used and what it cost, with no time, because you didn't wait for it. Answers from before October 19, 2026 have no receipt and keep their older line: 0 credits · 25312 tokens · first words 41s · done in 2m 05s — the credits it used, the tokens it read and wrote, how long you waited for its first words, and how long until it was done. The two times are what tell a slow answer from a stuck one — on Desktop AI your Mac can spend a minute reading a long question before it says anything. They're measured once, by ambientChat, and saved with the answer, so the same line is there when you reopen the conversation later, on your phone or the web, and in Q&A. That makes it a record you can compare over time — Desktop AI against the cloud models, or this month against last. Answers from before October 2026 show what they cost but have no times.

Hold an icon to find out what it is. Chat is a row of small glyphs and none of them is labelled, so press and hold any icon-only button and its name appears in a small white card. Let go without moving and nothing happens — it is a way of asking, not of doing. This works right across chat now: the round buttons above the message box, the save / print / export / copy icons under an answer, the context chips, the attachment thumbnails, the image cards, and the share buttons inside the Skills drawer and the share sheet. Only one name shows at a time.

(Two places are deliberately quiet. A button that already carries a visible text label has nothing to add, and the trash can on a document card doesn't answer a hold — that gesture picks the card up to drag it into a folder instead.)

#Web

The web interface at app.ambientchat.ai has the same chat, in your browser:

  1. Open the Chat page
  2. Type in the box at the bottom; Enter sends and Shift+Enter adds a new line (the desktop convention)
  3. Responses stream in above; while the AI is replying the ↑ send button turns into ■ Stop, as on the iPhone
  4. Attachments: drag a file onto the chat, or use the attach button, to send images or PDFs — or Pick from myAI Context to attach an existing item or document, which scrolls through the whole folder as described above. The same Just discuss / Save as document / Add to inventory choice appears, with a folder picker showing only folders of the matching type
  5. Ask AI from a detail view: open any item or document in myAI Context and click the teal Ask AI pill in the header — a fresh conversation opens with it attached, ready for your question
  6. Answers hand you the actual thing here too — the real card for one record, thumbnails for a few, an expanding "24 items" chip for many. Clicking one opens it, and the address bar updates so you can bookmark or share that exact record
  7. Model picker: click the model button (top-right) to switch provider/model — your choice is remembered on this device
  8. Past conversations: the Conversations column beside the chat has a panel button at its top-left — click it to hide the list, and click the same button (it stays in the same place, now at the top-left of the chat) to show it again. ⌘\ on a Mac, or Ctrl+\ elsewhere, does the same. The list starts hidden on a laptop-sized window and showing on a wide screen; once you open or close it, the web app remembers your choice on that browser. The + button next to it starts a new conversation, whether the list is showing or not. On a phone browser, ☰ opens the same list
  9. Second opinion: switch models and ask, or press ↻ with an empty box to put your last question to the model now selected (see Second Opinions)

Voice input and spoken replies are on the phone only — the web chat is text-only for now, not because a browser cannot speak but because we have not brought it over yet; a screen reader reads a web answer aloud either way. Icons name themselves on the web too: rest the mouse on one for half a second, or on a touchscreen — a phone browser, an iPad — press and hold it, and its name appears just above it. As on the phone, holding is a way of asking: letting go doesn't press the button. Every icon button on the web also has a screen-reader name, so the send, attach, skills, history and per-answer controls all announce what they are rather than just "button" — and every one of them is now a real button you can Tab to and press with Enter or Space, and that a screen reader lists when you ask it for the buttons on the page. That applies right across the web app, not only in chat.

#Ask without leaving the screen

While your AI is walking you through something on a screen, a small Ask about this screen button floats over it. Tap it and the coach bubble opens: a little chat right there, so you can ask "what goes in this box?" without leaving what you're doing. It is the same conversation as the Chat tab — your question and the answer are there too — and Open in Chat takes you to it. When the answer offers to take you somewhere or fill in a form, its Take me there or Fill in button is right under it in the bubble, too. The AI pill on a scan card (Radar, AI Lookup and AR) asks in the same bubble.

If you walk away from a walkthrough partway, the bubble offers to pick it up — "We were on step 2 — resume?" — with Resume and Not now. This works on the phone and on the web.

#Let your AI fill something in

Ask your AI to fill something in — "call it Car keys, in the Garage" or "share my list with Mathieu so he can edit" — and a Fill in button under the answer opens the form with your answer filled in and highlighted (Filled in by your AI. Nothing is saved until you save it.). Nothing changes until you save, or tap the suggested place, and Back brings you to the conversation. It works for an item's name and place, and for who a document is shared with and whether they can edit — on the phone and on the web.

#When your AI can't walk you through it

Not every model can walk you through a task on screen. If the one you picked can't, it may offer a stronger model from the same family for just that walkthrough: a line under the answer names that model, says it costs more and that afterwards you're back on your own, with a Walk me through it with … button. Tap it and that walkthrough runs on the stronger model, which costs more; afterwards you're back on your own model, and your model choice never changes. Ignore it and nothing changes. The same button shows under the answer in the coach bubble, so you can say yes without leaving the screen you're on. This works on the phone and on the web, but not with Desktop AI.

Don't want to be asked each time? Turn on Use a stronger model for walkthroughs automatically in Control → AI Model Selection (on the phone and the web; it's off until you turn it on), or just ask chat to "always use the stronger model for walkthroughs". Then the stronger model walks you through it straight away, with no question first, and a line under the answer names it — for example "Claude Opus 5.5 is walking you through this. Afterwards you're back on Claude Haiku 5.5." It costs what that stronger model costs, and the next ordinary question goes back to your own model. Turn it off, or ask chat to "ask me before using a stronger model", to get the question back. It never applies to Desktop AI.

#Copying an answer, or just part of one

On the iOS app, tap the copy icon under any answer — or press and hold the answer itself — to open the copy sheet. The same sheet opens from the copy icon on a Q&A card. It shows the answer as plain text, which is exactly what will land on your clipboard.

  • Blocks (how it opens) — the answer is broken into its natural pieces: each paragraph, each numbered or bulleted point, each heading, each code block. Tap the ones you want, then Copy selected. Tap again to untick.
  • Free text — switch to this tab to press and hold a word and drag the handles, the standard iOS way, when you want a phrase from the middle of a sentence.
  • Copy all takes the whole answer, from either tab.

On the web interface you don't need any of this — just click and drag across the answer the way you would on any web page, then ⌘C / Ctrl+C.

#Tips & Tricks

  • Tell it when something's gone. "We drank the Opus One", "used up the olive oil" or "sold the bike" marks the item Finished: it stops counting as something you have — the AI won't suggest a recipe with olive oil you've used up — but it keeps its photos, receipt and your 👍 or 👎. "I bought more olive oil" brings it back.
  • It knows your taste. Every answer can see what you gave a 👍 or 👎 — even things you've finished. A wine you loved and drank is something it will suggest at a restaurant, in a shop or as a gift, never for tonight's dinner (it isn't in the house); something you disliked it won't recommend. Ask "what do I like?" for the full list, or see it on the Me tab under Taste. It also always has a short summary of your taste, which you can read, edit or regenerate at the top of Me → Taste — and you can say "change my taste summary to…" in chat.
  • Tell it what you thought. "I loved that Opus One" or "never buying that coffee again" gives the item a 👍 or 👎, and the answer shows the item's card with its new mark. Cards for things you've reacted to always carry the 👍 or 👎, so you can see your taste at a glance.
  • Be specific with your questions. "What's expiring this week?" is better than "What about my food?" The AI has access to your entire inventory and document history, so the more precise you are, the more useful the answer.
  • The AI remembers your conversations. You can reference past chats: "What did we talk about last Tuesday?" or "Remind me what you said about my camera collection." It searches your conversation history automatically.
  • Try different models for different tasks. Some models are better at creative writing, others at precise analysis. If one model's answer isn't quite right, switch to another and ask again — your context carries over.
  • Press Return for a new line. On the iOS app, the Return key inserts a line break, so you can compose longer, structured prompts (steps, lists, several questions at once). Your message only sends when you tap the ↑ button. On the web interface it follows the desktop convention instead: Enter sends, Shift+Enter adds a new line.
  • Ask for a table — it reads on a phone. When an answer has a wide table (four or more columns, or three with paragraphs in them), the iPhone shows each row as its own card, every value labelled with its column heading, so nothing is squeezed into a sliver. On a wider screen it's a proper table that slides sideways when it's wider than the answer. Saved answers, notes and the web app follow the same rule, and so does a shared link.
  • Get a second opinion in one tap. Leaving the text box empty turns the send button into a ↻ repeat icon. Switch models first, then tap it to put your last question to the newly-selected model, which sees the earlier answer and which model wrote it. For an answer that isn't influenced by the first one, see Second Opinions → Get an independent answer.

#Example Prompts

Here are some things you can ask that take full advantage of your personal context:

  • "What's expiring this week?"
  • "How much is my electronics collection worth?"
  • "Create a shopping list from my low-stock pantry items."
  • "What did we talk about last Tuesday?"
  • "Show me all my receipts from Costco this month."
  • "What spices do I have that go with chicken?"
  • "Summarize the warranty for my laptop."
  • "What's in my medicine cabinet that expires soon?"

#Options

Setting What It Does Default
AI Model Which AI provider handles your questions Gemini
Voice Output Whether AI responses are read aloud Off
Voice Input Language Language for speech recognition Device language

#Every Answer Is Signed

ambientChat lets you use models from all three frontier labs, so every answer tells you exactly which one wrote it:

  • In the chat, each answer carries the provider's brand icon and the friendly model name (e.g. Gemini 4 Argon, Claude Opus 5.5) above the bubble — including answers in reopened conversations and each model's answer when you switch models mid-conversation. Answers that arrived from an external AI connected over MCP are labeled "via " instead.
  • In the Q&A view (myAI Context → Q&A), every saved answer card shows the same brand icon + model name, a chip for the surface it came from (Text, Voice, or MCP), when it happened, and — if the exchange was part of a named conversation — the conversation's name, so you can tell at a glance which chat it belongs to. Under each answer is the same receipt the chat showed — what it used, what it cost and how long it took, with the documents and the token count a tap away (an older answer shows its older line, tokens included) — so you can look back over weeks of answers. An answer another AI app saved into your history was never metered here, so it shows no cost at all rather than "0 credits".
  • The Q&A view is a scannable index: every conversation — single question or long thread — shows as one compact row with the stacked logos of every model that took part, its name, answer count, and time, so a screenful shows dozens of conversations. Everything starts collapsed; tap a row to expand its full answers in place (your expansions are remembered next time), and searching expands the matches automatically.
  • Q&A search reaches every conversation you have ever saved — questions, answers and conversation names — not just the ones the tab has loaded, and a line under the search box says how complete the answer is ("1 match", "200+ matches — scroll for more"). On iOS, offline, it can only look at what is already on the phone and the line says so. On the web the list also loads more as you scroll, so nothing older than the first hundred is out of reach.
  • The ↗ glyph on a row reopens that conversation in Chat so you can pick up where you left off.
  • An answer that made a picture shows it in Q&A too — the same Image Studio card you saw in Chat (or the side-by-side pair, where you can pick a favorite or keep both — and once you have, the pair shows your choice wherever you see it again, in Chat or Q&A), along with the answer's sources and the records it showed you. To ask for another take, reopen the conversation in Chat.

#An Answer Only Claims What Happened

Whichever model answers, ambientChat checks each sentence against what actually happened in that turn before you see it:

  • Links. A web link stays only if a web search or tool returned it in that turn, or it was already in your message or your documents. A made-up address loses its link (the words stay), or the sentence goes.
  • "Done." "Added", "ticked off", "moved" and "sent" are kept only when that change really succeeded. When it did not, that sentence is left out and "I didn't actually make that change." follows the rest of the answer, so you are never told something was saved when it wasn't. An answer that only tells you where something is, or tells a story, is never touched. An email that only opened as a draft for you to send is never called sent.
  • A stuck answer that starts repeating the same characters is cut off rather than filling your screen.
  • Unasked records. The AI won't create a new item, document or folder you didn't ask for while doing something else.

#When a Big Document Would Cost You

Your AI reaches for your own documents when a question looks related to one. Usually that is exactly what you want. Occasionally it would pull in something enormous — a scanned contract, a long report — and a large document is the most expensive part of an answer.

It now asks first. ambientChat knows how big your documents are before it opens them, so anything big enough to crowd out the rest of your context — around two or three pages of text, or a few pages of scanned images — is held back and offered to you rather than quietly added:

A large document was held back Scanned contract · 12 pages · ~24k tokens if included [ Include it ] [ Leave it out ]

The answer waits for you. The card says Waiting for your choice, and nothing is written until you tap Include it or Leave it out — the same answer then continues, with or without the document. You have about a minute; if you do nothing, it carries on without the document (the cheaper choice) and the AI is told it exists and was not read, so it will say so rather than pretending it looked.

Occasionally you will see the older wording, "A large document was used for this answer". That means the document was cheap enough to include without asking, and you are simply being shown what it cost. The buttons then offer Use summary only (keep the text, drop the scanned pages) or Answer without it.

#Your choice sticks for the rest of the conversation

That is the point: the same document would otherwise match your follow-up questions too, and you would be asked about it — and charged for it — over and over. A chip appears above the message box showing what you decided (e.g. Scanned contract · Not using · this chat).

From Claude Code, Claude Desktop or another AI agent the same document is held back, but there is no card to tap and nothing waits: the agent's answer arrives without the document and tells the agent it was held back, with the cost, so it can ask you and call again to include it. Each of those calls decides for itself — nothing sticks between them. See Second Opinions → Via MCP.

  • Tap the × on the chip to change your mind. The document goes back to normal from your next question onwards.
  • Starting a new chat clears it. The choice belongs to the conversation, not to the document.
  • Reopening an old conversation brings its choices back, on whichever device you reopen it — phone or web.

Nothing appears for ordinary documents. A receipt, a short note or a one-page manual flows straight through, as before.

#Documents are matched by meaning, not by matching words

Your AI finds relevant documents by comparing meaning, not by looking for shared words. That matters: a question about your health plan and a document about your health plan share obvious words, but so do a dozen documents that have nothing to do with your question.

Very occasionally that meaning-based search is unavailable. When it is, ambientChat will not fall back to guessing from word overlap — it would rather use no document than the wrong one. The answer will say that document search was unreliable, so you know to ask again or name the document you meant. It won't quietly answer as though it had checked.

#Known Limitations

  • The AI cannot take actions on your behalf yet — it can tell you what's expiring, but it can't order replacements or throw things away for you. (Probably for the best.)
  • Every model is in the picker on every plan — your plan sets how much you can use, not which model you can pick, so a second opinion from any lab is open to free users too. Free users have a daily allowance that refills at midnight. When it runs out, chat shows a card saying so and when it refills.
  • Very long conversations may truncate older context. If the AI seems to have forgotten something from much earlier in the chat, start a new conversation — it still has access to your full inventory and documents.
  • Image analysis in chat depends on the selected AI model's vision capabilities. Not all models handle images equally well.
  • An answer never stops on its own announcement. If the AI writes only "Let me check the Shopping folder…" or "Here's a closer look…" and then stops, it is asked once to carry on, and you get the content — usually the line itself stays on the working line, and if it had already reached the answer, the content follows it. If it still gives nothing, the answer ends with a short note saying it couldn't produce an answer, so a promise is never left hanging.
  • A very large request — filing a couple of hundred items into folders in one go, say — can run past the time one reply is allowed, or past the number of steps it may take. When that happens, any action the AI had already decided on still runs, and then it stops looking things up and writes its answer from what it found, saying what it couldn't get to. Only if that write-up comes back empty does the answer end with a short note saying which limit was hit and whether that last action ran or failed — or, if the reply was cut short before it got to an action, that the answer itself was cut off. A turn that ends with that note and changed nothing is not charged. A follow-up question starts fresh rather than resuming, though a change that ran has already been made.
  • A reply can stream for up to nine minutes. If your connection to it drops before it finishes — a network blip, the app in the background, a very long answer — the answer is not lost: it is saved on the server as it completes, and the chat fills the bubble in from your Q&A history within a couple of minutes on both the iOS app and the web. If nothing turns up in that time you'll see "The answer was cut off." — and for the next half hour, coming back to the app (or back to the browser tab) checks again, so an answer that finishes late replaces the note on its own. Asking something new in that conversation stops the check for the earlier answer; it is still in Q&A.
  • A message that never left your phone or browser — the connection dropped before it reached the server — now says so at once: "Your message didn't reach your Mac — tap to send again" (or the model's name, for a cloud model). When your connection comes back, the chat sends it again by itself, once; if that fails too, tap the card to send it. A message the server did receive is never answered twice — a resend finds the answer it already saved.
  • Switching a conversation to Desktop AI starts your Mac reading it straight away, while you type your next question — so its first answer after the switch no longer waits for the Mac to read the whole conversation first.
  • A slow start is not a dead one. With Private AI the Mac reads your whole question — and any pinned documents — before it writes a word, which can take a minute or more; the chat now keeps waiting through that instead of calling the answer interrupted.
  • Stop (■ on the send button, on iOS and the web) works at any point, even while your attachments are still uploading. It really stops: the AI stops writing, takes no further steps (an action it was already in the middle of, such as filing items, finishes so nothing is left half-done), and you are charged only for what it had done up to that point. Whatever had arrived stays on screen and is saved to Q&A exactly as you saw it; an answer stopped before its first word goes away and is not saved. Stopped before the question reached any AI, it says Stopped — not sent, with Send again to ask it after all. A stopped answer carries a small Stopped line under it: in the chat, in Q&A, when you reopen the conversation, when you print or export it, and wherever it is shared (in the app, by link or by email), so nobody reads it as a finished answer. You can ask your next question in the same conversation straight away.
  • If an attachment doesn't upload — a dropped connection, say — the message is not sent, on iOS and the web. Chat names the file that failed ("kitchen.jpg didn't upload.") and puts your message back in the text box with all its attachments, so you can send it again; the ones that did upload are not uploaded a second time. It never answers about some of your attachments as if it had seen them all.
  • If the chat can't start an answer at all — your Mac isn't reachable for Private AI, you're sending faster than allowed, the conversation is still working on your last question, or the question is too big for the model — it says why, on iOS and the web. Nothing was produced in that case, so there's nothing to look for in Q&A. If your Mac drops partway through a Desktop AI answer — after the AI has started looking something up — the answer ends with the same reason, such as "Your Mac isn't connected right now.", and that is what Q&A keeps.
  • A reply that is working without writing anything you can see yet — the AI writing out a long action such as moving hundreds of items, an image being generated (for up to five minutes), or the chat waiting for you to approve a change — keeps the chat waiting instead of giving up on it. Gemini models write an action all at once rather than a piece at a time, so a very long one can still look silent while it is being written. A reply that shows no sign of work for 90 seconds ends with "That one went quiet", and that note says you weren't charged only when the charge was actually undone.
  • After the AI looks something up — your inventory, documents, the web — it reads the result before it writes again. While it does, the answer shows Going through what it found… under what it has written so far, on iOS and the web. With Desktop AI on your own Mac that reading can take tens of seconds; the indicator means it is still working.
  • Some models think before they write a word — Gemini 4 Argon most of all — and send nothing while they do. Chat gives Argon up to three minutes to start each step, and if a step still hasn't started it asks once more by itself. If that fails too, the answer says so straight away — "Gemini 4 Argon took too long to respond, so this answer wasn't saved. Please try again." — on iOS and the web, instead of leaving Going through what it found… spinning. The same goes for any answer that fails before a word of it appears and before it changed anything: it ends with a short note to try again, rather than a spinner, because there is nothing to look for in Q&A. If part of the answer had already appeared, that part stays.

#Version History

Version Date What Changed
95 2026-11-13 Adds An Answer Only Claims What Happened: every model's answer keeps a web link only if that turn's search or tools returned it (or you supplied it), says "added", "ticked", "moved" or "sent" only when that change succeeded and otherwise leaves that sentence out and ends the answer with "I didn't actually make that change.", never calls a drafted email sent, cuts a stuck repeating answer, and won't create a record you didn't ask for AMB-3249
94 2026-11-11 Adds Don't want to be asked each time? to When your AI can't walk you through it: a switch in AI Model Selection (off by default, also settable by asking chat) runs the walkthrough on the stronger model without asking, names it in a line under the answer, bills at that model's price and goes back to your own model afterwards. Never with Desktop AI AMB-3213
93 2026-11-08 On the web, the button that shows and hides your conversation list now sits on the list itself — a panel icon at the top-left of the Conversations column, which stays in the same spot when the list is hidden, so one click undoes the other. It's called Conversations everywhere (no more "History"), + starts a new conversation in both states, and ⌘\ (Ctrl+) toggles it from the keyboard. AMB-3110
92 2026-11-08 The AI always has a short summary of your taste, and connected apps get it when they connect. AMB-3071
91 2026-11-07 On the web, your conversation list sits behind a History button at the top-left of the chat, giving the chat more room: hidden by default on a laptop-sized window, showing on a wide screen, and your choice is remembered. AMB-3085
90 2026-10-29 If you tap Approve or Deny on the lock screen and your phone puts ambientChat to sleep before the answer gets through, for example on a weak signal, ambientChat now sends it again the next time it opens. AMB-2835
89 2026-10-29 When you move several Q&As to Trash and only some of them can be moved, ambientChat now says both: for example "2 Q&As moved to Trash. 1 couldn't be moved." The ones left behind come back into your list. Before, the iPhone counted only the ones that moved, and the web said nothing while the others stayed out of sight until you reloaded. AMB-2833
88 2026-10-29 Approve and Deny on an approval notification now work even when ambientChat is closed. Before, an answer tapped on the lock screen of a closed app could be lost without a word, and the assistant waited and timed out. If the answer can't be sent, for example with no connection, a notification says so, and tapping it opens the chat where you can answer again. AMB-2747
87 2026-10-28 On the phone, tapping Fill in for an item under a chat answer now keeps you in Chat: when you're done, Back brings you to the conversation, as it already did for a document. AMB-2772
86 2026-10-28 When your AI offers a stronger model to walk you through something, the Walk me through it with … button now shows in the coach bubble too, on the phone and the web. AMB-2770
85 2026-10-28 In the coach bubble, an answer's Take me there and Fill in buttons now show right under it, on the phone and the web, even if you never opened Chat. AMB-2693
84 2026-10-28 Adds When your AI can't walk you through it: a model that can't coach may offer a stronger one from its own family for just that walkthrough. It costs more, you're back on your own model afterwards, and ignoring it changes nothing AMB-2584
83 2026-10-27 Adds the coach bubble: ask about the screen you are on without leaving it. It uses the same conversation as Chat, with Open in Chat, and offers Resume? if you walked away from a walkthrough AMB-2692
82 2026-10-27 Adds Ask to the item editor, the Last seen picker and the share sheet. It opens the coach over the form, and anything the coach fills in lands in that form, highlighted. Nothing changes until you save AMB-2673
81 2026-10-27 Every model is in the picker on every plan; your plan sets how much you can use. Free allowances refill daily, and a card says when AMB-2584
80 2026-10-27 Adds Fill in: ask your AI to fill in a name, a place or who a document is shared with, and a button opens the form with it filled in. Nothing changes until you save AMB-2584
79 2026-10-26 An Image Studio picture that runs unusually long now ends by telling you what happened to it: if it finished, "it's in your Documents" with an Open image button (you were charged once, for the picture you got); if it didn't, "you weren't charged — please try again". Only when neither can be known does it say the image couldn't be finished — on the phone and on the web alike (AMB-2564)
78 2026-10-26 On iOS, an answer that ends with an error finishes at once. When something went wrong after the answer had started, such as your Mac dropping during a Home AI answer, the iPhone kept the finished answer under a spinner for a few seconds while it looked it up again in Q&A; it now finishes straight away with what you already see, as the web does AMB-1872
77 2026-10-25 A help guide is never treated as one of your documents. A question like "What model am I talking to?" could pull a whole Don't Panic guide into the answer and stop to ask whether to include it, as if it were a private file of yours; guides now stay out of the documents matched to your question — the AI looks up help when it needs it — and your own documents are held back and offered exactly as before AMB-2526
76 2026-10-25 Every AI model knows ambientChat. Whichever model you pick — including Desktop AI on your Mac — now carries a short map of the app's features and where they run, and looks up the relevant help before answering a "how do I…" or "why isn't this working" question, instead of guessing at steps AMB-2526
75 2026-10-25 "How do I…" questions are answered from the right part of the guide. When you ask how to do something in ambientChat, the AI now reads the section of the help guide that covers it — the steps and the limits around them, not one stray sentence — and says which guide it came from AMB-2526
74 2026-10-24 Switching a conversation to Desktop AI gets your Mac reading it at once. On iOS and the web, picking Desktop AI in a conversation that already has answers starts the Mac reading it while you type, so the first answer after a switch from a cloud model starts sooner instead of waiting for the Mac to catch up on the whole conversation
73 2026-10-24 A message that never reached the server is sent again, not lost. On iOS and the web, a message whose connection dropped before it arrived says "Your message didn't reach your Mac — tap to send again" (or the model's name) straight away, instead of waiting two minutes and calling the answer interrupted. When the connection is back the chat resends it once by itself; the server never answers the same message twice
72 2026-10-24 An attachment that fails to upload stops the message, on iOS and the web. Chat names the file that did not upload and puts your message and its attachments back in the text box to send again — instead of answering about the ones that made it as if it had seen them all (AMB-2294)
71 2026-10-21 A question you stop before any AI was asked says "Stopped — not sent", with Send again, on iOS and the web (AMB-2291)
70 2026-10-21 An image now says when some of the reference photos you picked were not used — one image takes up to three — instead of leaving them out without a word, in Image Studio and in chat, single images and Dual pairs alike (AMB-2199)
69 2026-10-20 A side-by-side image pair remembers your pick. Once you pick a favorite or keep both — by tapping, or by telling the assistant "keep the left one" — the pair shows your choice right away, and again whenever you reopen the conversation or find it in Q&A, on iOS and the web. Tapping an old pair that was already decided shows the real choice, never "Both saved" when you kept just one AMB-2101
68 2026-10-19 On a touchscreen, press and hold a web icon to see its name. In a phone browser or on an iPad, holding any icon-only button shows what it is, just as resting the mouse does on a computer and holding does in the phone app — and letting go doesn't press it (AMB-2200)
67 2026-10-19 Ask chat how ambientChat works and get the answer. Help questions now search the Don't Panic guides for every account, not just the one that holds them, and the answer is the passage that matched — even a sentence deep inside a long guide, such as where album covers and track lists come from AMB-2123
66 2026-10-19 Q&A shows the same receipt as Chat. The line under a saved answer is now exactly the chat's — Used 1,387 items · 3 documents · Study · 0.42 credits · 9.8s, tap for the documents and tokens — on iOS and the web, instead of a rounded-up "This response used 1 AI Credit". An answer saved by another AI app shows no cost rather than "0 credits". The iOS Show AI tokens switch is gone: tokens are where Chat puts them, in Q&A too AMB-2099
65 2026-10-16 Q&A shows the pictures your answers made. A saved answer that created an image in Image Studio now shows its image card in Q&A, the same as in Chat — the side-by-side pair too — on iOS and the web AMB-2081
64 2026-10-11 How long an answer took is saved with it. The line under an answer — credits, tokens, first words and done in — now comes back when you reopen the conversation on iOS or the web, and Q&A shows the two times right after each answer's cost, so you can compare Desktop AI against the cloud models over time AMB-1961
63 2026-10-09 A stopped answer says "Stopped". An answer you stopped part-way used to look finished once you left the chat: in Q&A, in a reopened conversation, and to anyone you shared it with. It now carries a small Stopped line under it everywhere: the chat, Q&A, a reopened conversation, a print or export, and every kind of share, on iOS and the web AMB-1884
62 2026-10-07 Stop really stops — and the web's Stop is on the send button. Stopping an answer used to end it only on your screen: the AI finished it anyway, you were charged for all of it, the whole answer went into Q&A, and asking again in that conversation was refused until it was done. Now the AI stops when you tap ■, you're charged only for what it did, Q&A keeps exactly what you saw, and you can ask again at once. On the web the send button becomes ■ Stop while the AI answers, like the iPhone AMB-1870
61 2026-10-07 If your Mac drops partway through a Desktop AI answer, the chat says so. When the Mac disconnected after the AI had started looking something up, the answer read "Tool follow-up failed: 503" and then "something went wrong part way through"; it now shows "Your Mac isn't connected right now.", on iOS and the web, and that sentence is what Q&A keeps AMB-1682
61 2026-11-13 An answer only claims what happened — links must come from that turn's search or tools, and "added", "ticked", "moved" or "sent" only when that change succeeded; otherwise you read "I didn't actually make that change." (AMB-3249)
60 2026-11-13 Quick answers show sooner — an answer appears in the bubble from its first sentence instead of waiting on the working line; the AI's notes about what it will look up still stay on the working line (AMB-3055)
59 2026-11-13 An answer never stops on its own announcement — a reply that only says what it will do next ("Let me check…", "Here's a closer look…") is asked once to carry on, and you get the content; if nothing comes, a note says so (AMB-3144)
58 2026-11-13 The AI's "Let me check…" notes stay on the working line, not in the answer you read, hear, find in Q&A or export — on iPhone, the web and the coach bubble (AMB-3055)
57 2026-11-11 A long request ends in an answer, not just a note — when a reply reaches its time or step limit, the AI stops looking things up and writes its answer from what it found; the "hit its step limit" note shows only if that comes back empty (AMB-3112)
56 2026-11-09 Slow-thinking models get time to think — Gemini 4 Argon waits up to three minutes per step and asks once more if it has to; an answer that still fails says so straight away instead of spinning (AMB-3138)
55 2026-11-08 The AI knows your taste — likes and dislikes, finished favourites suggested when you're out, never for dinner at home (AMB-3068)
54 2026-11-08 Tell the AI when something's gone — "we drank the Opus One" marks it Finished; "bought more" brings it back (AMB-3067)
53 2026-11-07 Say what you thought of something and the AI gives it a 👍 or 👎; answer cards show the mark (AMB-3066)
52 2026-11-05 Anything an answer added or changed always gets a card, even when the answer doesn't mention it, so a record the AI made along the way can't go unnoticed (AMB-2985)
51 2026-10-21 A question you stop before any AI was asked says "Stopped — not sent", with Send again, on iOS and the web — instead of vanishing, or reading as an interrupted answer to look for in Q&A (AMB-2291)
50 2026-10-21 Watching it work, in everyday words: a lookup says what it's looking for ("Searching the web for …"), a long read on your Mac says roughly how long is left, a follow-up says your Mac is catching up on the conversation, and a cloud model that's working it out says "Thinking it through…" (AMB-2299)
49 2026-10-20 An answer that appears instantly from the Radar or Lookup AI button now shows its receipt — what it used and what it cost — and keeps it when you reopen the conversation (AMB-2175)
48 2026-10-20 The AI remembers every file you attached, not just photos and sheets. A PDF, Word or PowerPoint file or note you chose Just discuss for can be asked about again later in the same conversation, and a photo or PDF you chose Add to inventory for is remembered along with what the scan added (AMB-2045)
47 2026-10-20 A document opened from a chat answer, Scan or Sharing now goes < back to the Documents list, on iPhone and the web; the browser's Back still returns to the chat, and Q&A keeps its own Back (AMB-2218)
46 2026-10-19 iPhone: tapping the myAI Context tab now always takes you back to its start screen. It could leave you stuck on an item opened from chat, or stack a fresh, empty Inventory over the one you had a folder open in (AMB-2139)
45 2026-10-19 Desktop AI answers are shorter: each thing named once with what you asked about, the rest left to its card (AMB-2090)
44 2026-10-19 Desktop AI: an answer that names your things without linking them still shows their cards (AMB-2177)
43 2026-10-19 Watching it work: a line under a working answer says what it's doing — gathering your context (naming what), your Mac reading it with a progress bar, lookups, writing — instead of a spinner (AMB-2171). What an answer used: the receipt under each answer — Used 1,387 items · 3 documents · Study · on your Mac · 9.8s — tap it for the documents and the technical numbers; saved with the answer and shown in Q&A (AMB-2174)
42 2026-10-18 iPhone: an item opened from a chat answer or a new chat's "expiring soon" suggestion now has Inventory beneath it, so < goes to Inventory with its category bar instead of back to Chat, and you are never stuck on the item (AMB-2138)
41 2026-10-09 The Panel of Experts toggle is gone (AMB-1930). For a second opinion, switch models and ask, or press ↻ with an empty box — on the phone it works any time now, as on the web. See Second Opinions, which also explains how to get an independent answer
40 2026-10-06 Private AI is now called Home AI (AMB-1782)
39 2026-10-06 A slow answer keeps the chat waiting, a late one still arrives, and the iPhone has a Stop button. A Desktop AI answer with documents pinned could say "interrupted" while the Mac was still reading the question; the chat now waits through that. If an answer does finish after the note appears, coming back to the app or tab within half an hour fills it in. And on iOS the send button becomes ■ Stop while the AI answers, as the web's Stop button already did (AMB-1797)
38 2026-10-06 A document in an answer says which folder it's in, on iOS and the web, the way an item's card already did (AMB-1715)
37 2026-10-05 When the AI can't answer, the chat says why. If your Mac was offline for Desktop AI, or the chat refused the request (too many at once, the last question still running, a question too big for the model), iOS said the answer "was interrupted" and sent you to look for it in Q&A, where it never was; the web showed a generic error. Both now show the actual reason, such as "Your Mac isn't connected right now." (AMB-1766)
36 2026-10-02 An answer says "Reading the results…" while the AI reads what it looked up, on iOS and the web, instead of sitting still — most noticeable with Private AI, where that can take tens of seconds (AMB-1718)
35 2026-09-29 "Just discuss" means it for spreadsheets too. A spreadsheet attached in chat is no longer saved to Documents unless you chose Save as document or ask; the AI can still read every row and import them into your inventory for as long as the conversation lasts (until you trash it, or 30 days). Add to inventory on a spreadsheet now imports its rows (AMB-1608)
34 2026-09-28 Chat tells you when it saves an attachment to Documents — "Saved “albums.xlsx” to Documents" — on iOS and the web, and Documents shows the new file straight away; attaching the same spreadsheet again reuses the saved copy rather than making a duplicate (AMB-1595)
33 2026-09-28 A spreadsheet attached in chat is saved to Documents, whatever you choose. The AI used to read a "Just discuss" spreadsheet and then be unable to import it or read past its first rows, because it had no saved file to work from; now it does, on both surfaces (AMB-1590)
32 2026-09-25 Pick from myAI Context scrolls through a whole folder, on iOS and web. It loaded the first 200 documents or items and stopped without a word, so in a bigger folder the thing you wanted to attach could be missing from the list; it now loads a screen at a time as you scroll, to the end of the folder (AMB-1255)
31 2026-09-25 Reworded the Web note on voice: spoken replies are on the phone only for now, not a browser limitation — and a screen reader reads a web answer aloud regardless
30 2026-09-22 A long action, an image, or an approval no longer ends as "That one went quiet". The chat gave up on any reply that showed nothing for 90 seconds, even when the AI was writing out a long action, an image was generating, or the chat was waiting for your approval — and the answer then finished unseen. It now waits while the reply is working, and the "went quiet" note only says you weren't charged when the charge was actually undone (AMB-1279)
29 2026-09-18 A long answer streams for up to nine minutes, and a dropped connection no longer loses it. The reply window was five minutes and a longer turn simply went quiet on screen while the server finished it; now the window is nine, and if the connection still drops, the chat fills the answer in from your Q&A history on both iOS and web instead of leaving an empty bubble (AMB-1219)
28 2026-09-18 A long request no longer throws away its last action. When a reply runs past the time or step limit, the action it was about to take now runs before the turn stops, and the closing note says which limit was hit and whether that action ran or failed — it used to blame "tool steps" whatever the cause and promise a "continue" it could not keep
27 2026-09-17 Noted what happens from an AI agent (Claude Code, Claude Desktop): a large document is still held back, but the agent is told rather than shown a card, nothing waits, and nothing sticks between calls — with a link to the Panel of Experts page
26 2026-09-15 A held-back document now pauses the answer until you decide. Tapping Include it continues the same answer with the document in, instead of quietly failing while an answer without it arrived anyway; if you don't choose within about a minute it continues without
25 2026-09-15 Q&A search asks the server and reaches every saved conversation, with a completeness line under the box — iOS and web; web Q&A now loads more as you scroll
24 2026-09-13 Every control in the web app is now a real button: you can reach it with Tab, press it with Enter or Space, and a screen reader finds it when you ask for the page's buttons. They were named last release but not reachable — a mouse was the only way to press most of them
23 2026-09-13 Every icon button in the web app now announces what it is to a screen reader, and the sign-in form's fields say which is which — they used to announce "edit text, blank"; the Web section now also says that press-and-hold is iOS-only
22 2026-09-13 Press and hold any icon-only button in chat to see its name — it used to work only on the round buttons above the message box, so the gesture read as broken everywhere else
21 2026-08-21 Documents are no longer pulled into an answer on word overlap alone when meaning-based search is unavailable — the answer tells you search was unreliable instead of quietly using a document that merely shared some words
20 2026-08-21 The size at which a document gets held back now adapts to the AI model you have chosen, instead of a fixed figure that never triggered on the default model
19 2026-08-20 A big document is now held back and offered to you BEFORE it is used, instead of being added and billed first — with the page count and token cost shown up front, for every document type rather than just PDFs
18 2026-08-20 Choosing to drop a big document now sticks for the whole conversation — the note under the answer says what was applied instead of looking like nothing happened, and a chip above the message box shows the choice with an × to undo it
17 2026-08-19 The copy sheet now opens on Blocks — tap the paragraphs, bullets or code blocks you want and Copy selected; press-and-drag selection moved to a Free text tab beside it, so copying part of an answer no longer depends on the drag working
16 2026-08-14 Word, Excel and PowerPoint files can now be attached to chat directly and asked about — their text is extracted for you; an unreadable one says so instead of being answered as if it had been read
15 2026-08-14 Named the file types chat actually accepts (it used to say "or other documents", which read as though Word files worked); a file chat can't take now says so instead of disappearing
13 2026-08-08 The mic joined the button row above the message box (it used to vanish as soon as you typed); the model button now shows the price band as $–$$$$ instead of the band name
12 2026-08-05 "Pick from My Stuff" renamed Pick from myAI Context — My Stuff was never a real surface; the picker at Home now shows only Home's documents instead of every document in the library
1 2026-03-01 Initial guide
11 2026-07-26 Attachments now take ONE choice — Just discuss / Save as document / Add to inventory — and the folder picker shows only folders that can hold what you're saving; both surfaces
11 2026-07-26 Attaching a photo also works for items you already own — adds or replaces that item's picture
10 2026-07-25 Attach a photo and ask the AI to save it — the item keeps your photo as its image; see Inventory v10
9 2026-07-23 Corrected Panel mode description — it's sequential (switch model, then ↻ re-ask) not "ask several models at once"; see Panel of Experts v3
8 2026-07-20 Q&A is now a collapsed-by-default index — every conversation is one row, tap to expand, ↗ opens it in Chat; search auto-expands matches — both surfaces
7 2026-07-19 Q&A conversation collapse choices are now remembered across visits — both surfaces
6 2026-07-18 Q&A view groups answers by conversation (stacked model logos, collapsible); conversation names are tappable and reopen the thread in Chat — both surfaces
5 2026-07-16 New "Every Answer Is Signed" section — model brand icon + friendly name on every answer (chat + Q&A view), surface chip and conversation name on Q&A cards
4 2026-07-16 "Ask AI about this" from item/document detail on iOS and web — fresh thread with the asset attached, you frame the question; noted Pick from myAI Context in both attachment sections
3 2026-06-07 Added Web section — chat in the web interface (Enter to send, model picker, Panel, attachments); fixed "web console" wording
2 2026-06-06 iOS: Return now adds a new line; tap the ↑ button to send. Documented the web Enter / Shift+Enter convention and the ↻ re-ask-the-panel shortcut in Panel Discussion mode.