teach-learn MCP

One profile of you that every assistant can fill in and read. Teach it once from the assistant that knows you best; every other assistant then starts already knowing you.

teach

An assistant tells it what it knows about you. Each response asks follow-up questions, so the assistant keeps going on its own. When it runs out, the questions go to you.

learn

Before helping you, an assistant asks what it should know for the task. It gets a briefing, what to ask next, and what only you can answer.

The server steers the conversation through its responses (next_step, status, follow-up questions), so it works in any MCP client without special setup.

Quick start

Get your connector URL
Use the private connector URL you were given. It ends in /mcp/ followed by a long token. Whoever runs this server can open /setup/<token> to get this page with the URL filled in.
Connect your assistant
1. Settings → Apps & Connectors → Advanced → turn on Developer mode.
2. Create a connector named teach-learn with your URL. Authentication: None.
3. In a chat, enable teach-learn from the tools menu.
1. Settings → Connectors → Add custom connector.
2. Name it teach-learn and paste your URL. Leave the OAuth fields empty.
3. In a chat, enable it from the tools menu.
Run once, then start a new session:

claude mcp add teach-learn --scope user --transport http YOUR_CONNECTOR_URL

Add to ~/.cursor/mcp.json:

{ "mcpServers": { "teach-learn": { "url": "YOUR_CONNECTOR_URL" } } }
Teach it about you Paste the teach prompt into the assistant that knows you best (usually ChatGPT). Expect several rounds of teach calls, then a few questions for you.

Prompts

Teach it about you

Use the teach-learn tool. Teach it everything you know about me, from your memory and our past conversations: who I am, my work, my projects, the people I work with, my preferences, and how I like to communicate. Call teach, then follow each response's next_step: answer its follow-up questions yourself whenever you can, and only ask me the ones you can't. Keep going until its status is "complete", then tell me what you taught it.

Teach from another assistant

Use the teach-learn tool. Call learn first to see what it already knows about me, then teach it anything you know that's missing or out of date. Follow each response's next_step until its status is "complete".

Before a task

Before you start, call learn on the teach-learn tool with purpose set to what I'm asking for. Use what it tells you, follow its ask_me_next suggestions if you need more, and ask me its ask_the_user questions only if they matter for this task. Teach back anything new I tell you.

Make it automatic

For ChatGPT custom instructions, a Claude project's instructions, or a CLAUDE.md.

You have a teach-learn tool that holds my profile. At the start of any task about me or my work, call learn with the task as the purpose before answering. Whenever you learn something durable about me (a new project, a preference, a change in role), call teach with it.

Keep it current

Teach the teach-learn tool what's changed:

Tool: `teach`

Adds what the calling assistant knows about the user to the profile, and returns what to send next. Not read-only: it writes facts and questions.

Input

Field	Type	Description
`content` required	string	What the assistant knows about the user, in plain prose. Can cover any number of topics at once.
`answers_to`	integer[]	Ids of `next_questions` from an earlier response that this `content` answers. Lets the server close them.

Response

Field	Type	Description
`status`	string	What the assistant should do next: `continue`, `ask_user` or `complete`. See status.
`next_step`	string	The instruction for the assistant, in words. Follow it.
`next_questions`	object[]	Questions the assistant can probably answer itself: `{id, topic, question}`. Answer them in the next `teach` and pass their ids in `answers_to`.
`ask_user`	string[]	Up to 3 questions only the user can answer, worded for them. Filled when `status` is `ask_user`.
`conflicts`	object[]	New information that contradicts a stored fact: `{fact_id, existing, incoming, question}`. Nothing is overwritten until it's settled.
`learned`	string	One line on what was taken in from this call.
`stored`	integer	New facts stored by this call.
`replaced`	integer	Older facts this call replaced with newer information.
`profile`	string	A one-line summary: total facts and which topics are well covered.
`coverage`	object	Each topic mapped to a coverage level.

Example response

{
  "learned": "Took in her role, the KYC project, team, tools and style.",
  "stored": 16,
  "replaced": 0,
  "profile": "16 facts; well covered: work, people, tools",
  "coverage": { "identity": "thin", "work": "good", "projects": "thin", "goals": "none", … },
  "conflicts": [],
  "next_questions": [
    { "id": 1, "topic": "projects", "question": "What are the main open risks on the KYC automation project?" },
    { "id": 2, "topic": "preferences", "question": "What tasks does she most often ask you for help with?" }
  ],
  "ask_user": [],
  "status": "continue",
  "next_step": "Call `teach` again now, without waiting for the user: answer as many of next_questions as you can…"
}

Tool: `learn`

Answers from the profile, shaped to what the assistant is about to do. Read-only.

Input

Field	Type	Description
`question`	string	A specific question about the user. Leave empty for a general briefing.
`purpose`	string	What the assistant is about to do ("draft her weekly team update"). The answer is tailored to it.

Response

Field	Type	Description
`answer`	string	What to know for this question or purpose. It says what isn't known, and marks inferences as inferences.
`based_on`	string[]	The stored facts the answer rests on.
`ask_me_next`	string[]	2–4 follow-up questions to send back to `learn`, all answerable from the profile.
`ask_the_user`	string[]	What the profile is missing for this purpose that only the user can answer. Ask them, then `teach` the answers.
`next_step`	string	The instruction for the assistant.

On an empty profile, only answer and next_step come back. The next step says how to get the profile started.

Example response

{
  "answer": "Maya is a senior PM at Tidewell who owns merchant onboarding… Keep it short and bulleted. Don't write compliance wording for her. Not known: this week's progress.",
  "based_on": ["Senior PM at Tidewell", "Prefers short bullet answers", "…"],
  "ask_me_next": ["What goes into her Thursday update to Dana?"],
  "ask_the_user": ["Who receives the team update: just Omer's team, or Dana too?"],
  "next_step": "If your task needs more, call `learn` again with one of ask_me_next…"
}

Definitions

status

Value	Meaning	What the assistant does
`continue`	There are follow-up questions the assistant can probably answer.	Calls `teach` again right away with answers, without waiting for the user.
`ask_user`	The assistant has run out, but gaps remain that only the user can fill.	Asks the user the `ask_user` questions, then teaches their answers.
`complete`	Every topic is well covered.	Stops and tells the user what it taught.

Topics

Every fact and question belongs to one of these:

identityworkprojectspeopleexpertisetoolscommunication_stylepreferencesgoalsroutinesboundariesinterests

Fact

One durable claim about the user, in the third person ("Works at Tidewell as a senior PM"). Each fact records:

Property	Values	Meaning
confidence	`stated` · `inferred`	stated: the user said or did it. inferred: the assistant's own reading of them.
status	`active` · `superseded`	A superseded fact was replaced by newer information. It's kept for history, but never used in answers.
source	assistant name	Which assistant taught it (ChatGPT, Claude Code…), taken from the MCP client's name.

Coverage level

Value	Meaning
`none`	Nothing known about this topic yet.
`thin`	Something is known, but not enough to act on confidently.
`good`	Well enough covered that more questions add little.

Question

A gap the server wants filled. Each question has a topic, the reason for asking, and a status:

Status	Meaning
`open`	Asked and waiting for an answer. The server avoids asking it again.
`answered`	Answered in a later `teach` call.
`unanswerable`	The assistant said it didn't know. If it matters, it gets reworded for the user as an `ask_user` question.

Server behaviour

Transport: MCP over Streamable HTTP, one session per connected assistant.
First contact: an assistant connecting for the first time gets more, and broader, follow-up questions.
Deduplication: it doesn't store a fact it already has. Newer information replaces older facts (marked superseded).
Contradictions: raised as conflicts with a question that would settle them. Nothing is overwritten by a guess.
Errors: if the model behind the server fails or declines, the tool returns an MCP error and stores nothing from that call.
Restarts: sessions live in memory. After a restart, clients get a 404 and reconnect on their own, as the MCP spec expects.

Privacy & limits

The URL is the password. Anyone who has it can read and change the profile. Don't paste it into shared docs or chats.

It won't store secrets, credentials, account numbers or health details, even if an assistant offers them.
Text from assistants is treated as data about you, never as instructions to the server.
One profile per deployment: everyone who uses the same URL shares it.
Facts can't yet be edited or deleted from here.

teach-learn MCP

teach

learn

Quick start

Prompts

Teach it about you

Teach from another assistant

Before a task

Make it automatic

Keep it current

Tool: teach

Input

Response

Example response

Tool: learn

Input

Response

Example response

Definitions

status

Topics

Fact

Coverage level

Question

Server behaviour

Privacy & limits

Tool: `teach`

Tool: `learn`