memexdocs

Reference

How connections work

The connection guides, signing in from an assistant, the first prompt, checking a new connection, permissions, tokens, and managing connections.

Every connection speaks MCP, the standard assistants use to reach tools, at one address:

https://memex.tools/mcp

The connection guides

Connect an assistant (Settings › Assistants, also in the account menu, and above your name in the sidebar until an assistant has connected) opens the connection guides: four tabs, ChatGPT, Claude, Gemini Spark and Other, and for the first three every step as one numbered list: where to go in the assistant, the address to paste, approve sign-in with Memex, and the prompt to paste last.

Claude’s guide starts with an Add to Claude button that opens Claude’s Add custom connector form with Memex and its address filled in, keeps the steps by hand folded under it in case the button does not work, and has a step that allows all of Memex’s tools. Each guide is a few minutes. ChatGPT’s Developer mode and custom plugins and Claude’s custom connectors are available on free accounts, though a free Claude account holds one custom connector; Gemini Spark’s custom apps need a paid Google AI Pro or Ultra plan.

Signing in from the assistant

When you add the connector, the assistant sends you to memex to approve it. The consent page names the assistant and the site it connects from, shows which account you are signed in as, and says what the connection will be able to do: its writes land in your review inbox as pending, and nothing becomes verified without your approval. Connect it creates the connection. It appears in the Connected table at once, marked OAuth.

The prompt

The last step of every guide is a short prompt to paste into a new chat. It tells the assistant to use your memex as the source of truth about you, your work and your plans, to check it before answering about those things, to offer to save what you decide, and to read the memex docs first. For ChatGPT and Gemini the prompt asks the assistant to keep this in its own memory; Claude does not keep it between chats, so paste the same text into your Claude profile preferences or a Project’s instructions. See Your first message.

Checking a new connection

Ask the assistant to load the memex-docs skill and run the check below. It exercises read, search, instruction loading and write in that order, takes under a minute, and finds the failure you cannot see from the browser: a connection that reads but cannot write, or writes somewhere you never look. If you are an assistant that has just been connected, run it and report each result in plain language.

  1. Call health. Report the knowledge base name and the role it says you hold. If this fails, nothing below will work and the problem is the connection.
  2. Call search with no query. A new knowledge base holds the welcome notes memex wrote; any note will do. Fetch one with get and confirm you can see its body, which proves full bodies reach you, not titles.
  3. Call list_skills. memex-docs should be in it, along with the owner’s own skills. Load one with get_skill.
  4. Call propose with a short note of your own, titled Connection check, saying which assistant you are, when you ran this, and what the three steps above returned.
  5. Read the answer propose gave you. If it says the note is pending, tell the owner it is waiting in their review inbox and that this is correct: the review gate is doing its job. If it applied immediately, tell them this connection holds the curator role and writes without waiting.

Then say plainly which of the five worked.

Per-call permissions

Out of the box, ChatGPT and Claude ask you before each call until you tell them not to: in ChatGPT, Settings › Plugins › Memex, then Allow all actions; in Claude, Settings › Connectors › Memex, then Always allow for both Read-only tools and Write/delete tools, which Claude’s guide includes as a step with a link to its connector settings. Gemini asks before write actions and has no switch.

That is the assistant’s behaviour, not memex’s, and it is safe to switch off: with the agent role, new notes arrive in your review inbox and edits, deletions and merges wait for your approval; a curator connection’s creates and edits apply at once, and its deletes and merges still wait. See Stop the assistant asking every time.

Other: tokens

Other is the tab for anything that cannot complete a browser sign-in: a coding agent, a self-hosted agent, a script. Give the connection a name and Create token, then copy the token from the dialog; Copy token copies it again later. The agent sends it on every request as the header Authorization: Bearer <token>. Below the address, setups for Claude Code, Codex, Cursor and Gemini CLI give the command to run or the lines to add to the assistant’s settings file, with YOUR_TOKEN where the token goes. These connections show Manual in the table.

Managing connections

Connected assistants & agents lists every connection with its name, how it authenticated, when it was created and when it last called in. Its Actions hold Edit connection, Copy token for a connection made on Other, and Revoke.

Edit connection opens the identity editor: the displayed name, an optional description, and an icon (a provider logo, a glyph, or an uploaded image). The name and icon you choose are what every byline shows for that assistant across memex. For a connection made on Other it also shows the token, masked, with a copy button. memex keeps those tokens encrypted; one made before it kept them cannot be copied, and Make a new token replaces it, at once, so the assistant using the old one needs the new one.

Revoke marks the connection; Save changes confirms, and the connection stops working immediately. A revoked assistant gets “Invalid or revoked token” on its next call and nothing else. Its past bylines stay.

Every new connection, however made, starts with the agent role. See How roles work.