Skip to content

Let a model save when asked#

You can ask a model to keep something — "save this decision to the handbook", "remember that the launch moved to Friday" — and it proposes the note itself. Two ways, one rule: the model proposes, you save. What it proposes waits as a draft on the base's machine; you see its title and text, then save or discard it. A draft nobody decides on is dropped after a day.

Where you talk to the model The tool Where you confirm
A deployment of yours: the app's chat, the Playground, the hosted gateway save_to_knowledge A card in the chat: Save or Discard
An AI assistant connected to Astralyx (Claude, ChatGPT, Cursor…) save_to_knowledge_base The console's confirmation page, from the link the assistant shows

Before you begin#

  • A knowledge base to save into, in the deployment's workspace.
  • Permission to add to it (knowledge-bases:write-content) for whoever saves: the person who confirms.

A deployment's save_to_knowledge tool#

  1. Add the tool to the deployment's server-side tools (Give a model tools), naming the base:

    In spec.tools of the deployment (PUT /v1/eos/deployments/chat, the spec whole):

    "tools": [{"builtin": "save_to_knowledge", "knowledge_base": "handbook"}]

    The base must be of the deployment's own workspace. The model is told to use it only when the person asks to save, keep or remember something, and that nothing is saved until the person confirms. description and name change what it is told and called, as for any tool; approval is not needed — saving always waits for you.

  2. Chat with the deployment in the app or the Playground and ask it to save something: "Save the decision we just made to the handbook."

  3. The model calls the tool with a title and the text. The chat shows a card: the base, the title and the text it proposes.
  4. Choose Save (it becomes a note in the base, searched a few seconds later) or Discard. The model is told the note waits for you, never that it was saved.

On the hosted gateway, the tool's result carries the draft's id ("status": "awaiting_confirmation", knowledge_base, draft); your application shows it and saves it with POST /v1/metis/knowledge-bases/<base>/drafts/<draft>/commit, or discards it with DELETE …/drafts/<draft>. A machine's own gateway (an edge machine) does not propose notes: it answers the tool as unavailable, and the model says so.

Your AI assistant's save_to_knowledge_base#

An assistant connected over MCP (Connect your AI assistant) has save_to_knowledge_base in the metis toolset. It needs a connection that may write.

  1. Ask it: "Save a summary of this thread to the handbook knowledge base."
  2. It shows you the title and text it proposes, and a link to the console.
  3. Open the link: the console's confirmation page shows the base, the title and the full text, read from the base's machine. Choose Approve to save it, or Deny: the draft is discarded.
  4. The assistant learns the outcome with get_status.

Approving takes your browser session in the console: the assistant cannot approve its own proposals, and an API token cannot either.

Why it never saves silently#

A knowledge base is what later answers rely on. A note saved by a model on its own word — from a misunderstanding, or from instructions hidden in a document it read — would be cited as truth. So saving takes a person every time, and what the person approves is exactly what was proposed.