AI Assistant 2.0

The AI assistant is a chat that works with your libraries: it finds and counts entries, answers questions about your data, adds and changes entries, designs libraries, draws charts and pictures, and writes scripts and SQL.

Two versions exist: 2.0 (the chat described here, the default) and 1.0 (the earlier assistant, see Assistant 1.0). Switch in Settings → AI assistant version; turning it off there hides the assistant everywhere.

Opening the chat

Tap AI Assistant (the sparkles icon) wherever you see it: on the library list, in a library, on an entry card, in the editors. An app has its own Builder chat that builds and changes it, see Apps.

The chat opens as a panel over the screen. The place it was opened from only gives it context — the library, the entry or the script you are looking at — so it can answer “this” without being told which one. What the chat can do does not depend on that place: from any of them it can create, change and manage libraries and entries and look across all libraries you can open. To talk about a particular library or entry, it is handy to open the chat there, but not required; otherwise just name it.

The exception is code. Triggers, scripts and SQL are written in the editor they belong to, so open the chat from that script editor, script widget’s settings or the SQL Explorer: the chat is then bound to that script or query and can edit it. Likewise the app builder changes an app only when opened from that app.

An empty chat shows what it is best at in that place and suggestions to start with.

What it can do

  • Find and analyze: search entries by words, filter and count with SQL, compare, sum by month or category, read photos and files attached to entries.
  • Show: live entry cards you tap to open, charts (bar, line, pie…) you open full size and share, links to entries and libraries.
  • Add and change entries: create entries from a description, a photo, a document or a table; fill in or fix fields of existing entries.
  • Design libraries: create one or several linked libraries from a description; add, rename, move or change fields. Removing a field, changing its type or removing choice items asks for your confirmation first.
  • Draw pictures with an image model and put them into image fields.
  • Write scripts and SQL in the editor it was opened from; explain Memento’s script functions.
  • Answer questions about Memento from this documentation.
  • Use Web search for external or current information with Memento Cloud.
  • Read web pages, text, PDF, Word, Excel, PowerPoint and images by link.
  • Ask you when something is unclear, with options to tap.

It can move entries to the trash after you confirm the list. It cannot run scripts, open locked (password-protected) libraries or libraries your cloud rights hide, or change app settings — it tells you where to do these yourself.

Changes are applied at once and show in the chat as rows (field: old → new). Undo them by editing the entry, or restore an entry from Version history (cloud libraries). Deleted entries are restored from the trash.

Chats and history

  • Each place keeps its chat: the library’s, an entry’s, a script’s. New chat starts another one there.
  • Chat history lists all chats; filter by All, the current library or This entry. Swipe a chat away to delete it.
  • An answer can take a while; it keeps running in the background with a notification, and Stop ends it.
  • Long-press a message to copy it.
  • Chats are stored on this device only; they are not synced and don’t depend on the Memento account.

Attachments

Attach → Camera, Gallery or File: photos, PDFs, documents, tables. The assistant reads them, for example to create entries from a receipt, a price list or a spreadsheet. Some models do not accept photos; the model picker marks which do.

Models and AI credits

Settings → AI Models chooses the source:

  • Memento Cloud — models provided by Memento, paid with AI credits. Sign in to your Memento account to use it. Plans include monthly credits; the balance is on the Memento account screen and in the chat’s statistics. When credits run out, the chat says so and offers to upgrade.
  • OpenAI API — your own key for any OpenAI-compatible provider (OpenAI, OpenRouter, a local server…): the address, the key and the models to offer. Usage is billed by that provider.

The model of a chat is chosen in the chat’s model picker (Model; Manage models… for the list).

With assistant 2.0 and Memento Cloud, web search is enabled by default. To change the service or turn search Off, open Settings → AI Models → Web search and save your choice. Each service shows its price in AI credits per search; a search with no results also uses credits. The conversational model’s use of source text is billed separately as usual.

The assistant can search when an answer needs external or current information, in any chat, editor or app builder. Each search appears as a row with its query, service and source count. Tap it to see source links and text; tap a source to open it in your browser. Results contain short snippets or extracted passages. Answers can link to the returned sources.

Send a link and ask a question: “Find the warranty terms in this PDF” or “Describe the style of this picture”. Assistant 2.0 reads public web pages, text, PDF, Word, Excel and PowerPoint files, and images. You do not need to attach or save the file first. This works with either model source, even when web search is turned off.

Pages and documents are read as text. For long text, the assistant can read further portions of the same document. Pictures require a model that accepts images; the assistant reports when the selected model cannot read one. It keeps your selected model. Scanned or encrypted PDFs and pages that require sign-in or JavaScript may be unreadable.

Each read appears in the chat. Tap its row for the source link, loading time, text or image preview. Resources are limited to 10 MiB and five loading attempts per answer. Reading further portions of a retained document does not spend another loading attempt. Large documents are read in portions, up to about a million characters of text. Stop cancels the current load. Reading a link has no separate search charge; processing its text or picture uses the selected model as usual.

Instructions

  • Settings → Instructions → Your instructions — rules the assistant follows in every chat: “answer briefly”, “dates as DD.MM.YYYY”, “I am a vet, my entries are about pets”.
  • Library instructions — rules for chats in one library, its entries and scripts: Menu → Edit library → Main → Customize AI Assistant. Everyone who works with the library shares them; only those who may change the library’s settings can edit them.

Memory

The assistant keeps notes across chats: who you are, how you like it to work, what your libraries and fields mean. It writes them itself as it learns, and each change shows in the chat. Settings → Memory shows the notes: Edit them, Clear memory, or turn Use memory off (then it neither reads nor writes them). Memory stays on this device, is sent to the model with every request and holds up to 10,000 characters. Ask it to remember or forget something in plain words.

Privacy

What the assistant reads (entries, files, your messages, memory, instructions) is sent to the AI provider of the chosen source to produce the answer. Don’t use it with data you may not share with that provider. With web search enabled, the search query and any explicit language/country options also go to the selected search service. Conversation history and attachments are not automatically sent to it; a query can still include information the assistant has read. Source text is sent to the conversational model to write the answer. When reading a link, your device requests that URL directly from the website. The downloaded text or prepared image is sent to the selected AI model; conversation history and app credentials are not sent to the website.

To keep a library away from the assistant, set AI access in the library’s settings, or for all libraries in Settings → AI assistant → Library access:

  • Full
  • Read only — it reads entries and the structure but changes neither, nor the library’s scripts
  • No access — the assistant does not see the library at all

Was this article helpful?

Yes No

Related Articles