Skills and memories, governed

Teach it once. Every assistant knows it.

Your skills and memories live in one vault you own. Every assistant you connect reads the same current context, so nothing you taught one assistant goes stale in another. And anything they want to save waits for your approval first.

Your skills, in one place, used by many. How you write, how you plan, how you review: write it once and every connected assistant runs the same version.
Tested before deployed. Draft a skill, try it in the sandbox, deploy it when it behaves. Full version history, restore in one click.
Memories you scope. Store a fact once and choose which assistants can read it. Work stays work, personal stays personal.
Nothing lands without you. Every proposed change waits in your inbox as a diff: what leaves struck out, what arrives marked in.

Your vault

Verify your email once, and your vault opens.

What using it feels like

Everything funnels through your approval.

How it works

Three steps, and the third one is the point.

1Connect your assistants

Each assistant gets its own revocable link, and anything that takes a connector can hold one. One connection each, cut off any time, one at a time.

2They read one context

Your profile, memories, skills, and projects live in one governed store. Every assistant reads the same context, scoped by what you let each one see, so you stop re-explaining yourself.

3You approve every change

When an assistant wants to save something, the proposal waits in your inbox as a redline: what leaves struck out, what arrives marked in. Nothing lands without you.

How your vault thinks

Two surfaces: you, and how you work.

You

Who you are

The base layer. Every assistant you connect reads this first.

Marketing lead at a fast-growing company
Warm and direct, never corporate
Wants the answer first, the backstory only if asked

Read first byChatCodeResearch

You

What your assistants know

Facts and preferences, on the same page as who you are.

Standups are 15 minutes, no slides
The fall launch goes live October 1From a chat last week
My manager reads bullets, not paragraphs

Skills

How your assistants do things

A library of repeatable procedures, drawn from everything on your You page.

Draft email in my voice

ToJack

SubjectQuick favor

Thanks so much!

Turn my rough notes into a clean draft
Write the weekly status update

You approve every layer before it saves.

Control

And it is always yours

Teaching your assistants is easy. Making them forget is just as easy.

The old address on Maple StreetForgetForgotten

Chat✓Code✓Research✓

Gone from every assistant you connect, and on the record in your history.

Approve what an assistant wants to save before it lands
See which assistant changed what, and when

Contact

Building this for a team?

OriginKeep is invite-shaped while it grows. If your team needs governed context for its assistants, or you just have a question, write to us and a person answers.

contact@originkeep.com

Connections

Each assistant gets its own private link. Narrow what one sees, or revoke it, without touching the rest.

Review every write before it landsOff

Your connections

Loading connections…

Prime the assistant to use it

Add this once to the assistant's custom instructions so it loads your context at the start of every conversation.

At the start of each task, load my context from the OriginKeep connector: call get_profile and list_skills, and search it before assuming anything about how I work. When I tell you a durable fact, preference, or decision about me, call update_memory to save it so it is there next time.

Setup steps

Claude

Open settings

Settings → Connectors → Add custom connector.
Paste your connector link, save, and enable it.
Add the priming instruction to your Project instructions or profile.

Steps verified 26 Jul 2026 · desktop and mobile web

ChatGPT

Open ChatGPT

Use chatgpt.com in a browser. The Mac desktop app has its own Plugins, Apps, MCPs and Skills tabs, and nothing added there produces a working connection.
Settings → Plugins → Add → Create plugin.
Icon: upload a PNG of at least 256 by 256, under 10 KB. Download this vault’s icon, or use your own.
Name: OriginKeep. Description takes a sentence about what this is, not the link. You can paste: Your governed memory vault, holding your profile, memories, skills and projects, with approval before anything is written.
Connection: keep the toggle on Server URL rather than Tunnel, and paste your connector link there. A link pasted into Description instead leaves the plugin looking installed and doing nothing.
Authentication: None, not OAuth. The token inside the link is the credential, so there is nothing to sign in to, and OAuth will fail.
Tick the risk acknowledgement, then Create.
In each chat, type @ and pick OriginKeep from the Plugins group. Until you do, the plugin is installed but idle, and the message ChatGPT shows makes it look as though the plugin is missing.
Paste the priming instruction into Custom Instructions.

Steps verified 11 Aug 2026 on chatgpt.com in a browser. Whether a plugin created here also reaches the phone and desktop apps without being added again is something we have not tested.

Gemini

Open Gemini

Settings → Connected apps → Add a custom app, and paste your connector link as the MCP server URL. It files under Custom apps for Spark; a Last sync time means this vault answered.
Custom apps work in Spark conversations, not plain Gemini chat: type @ and pick this vault the first time you use it (personal Google account, Keep Activity on). After the first use Gemini reaches for it on its own.
Confirm the first real read under Activity → Reads here. Until a read appears, anything Gemini says it knows about you is its own memory, not this vault.

Steps verified 7 Aug 2026 · web + Spark. Gemini CLI, Vertex and ADK still take the link as a remote MCP server (mcpServers).

Grok

Open Grok

In Grok, add the link as an MCP connector or integration. For the Grok CLI, add the server to your grok config.
Enable it for your chats.
Paste the priming instruction into Grok's custom instructions.

Steps verified 26 Jul 2026 · desktop web and CLI

Microsoft Copilot

In Copilot Studio, add the link as an MCP server (custom connector).
Enable it for the agents or chats that should use it.
Add the priming instruction to the agent's instructions.

Any other MCP client

Cursor, Perplexity, and other MCP clients all take the same link.
Add it as a remote MCP server (Streamable HTTP transport).
Use the same priming instruction where the client allows custom prompts.

Skills

Reusable instructions your assistants follow. Search them, file them, pin the ones that should always apply, and walk each one from draft to deployed.

Add to your vault

Popular customizations. Tap one to see what it does, then add it.

already in your vault

v0.220.0 · 7341815

You're signed in

This vault is private. Yours would work the same way.

You are not seeing any of this vault's contents, and you are not meant to. A vault's profile, skills, and history stay private to its owner and the assistants they connect. That privacy is the whole product.

Want your own? Each vault is self-hosted end to end: your own store, your own assistants, nothing shared by default.
Invited to collaborate? The vault owner sends you a setup link. Open sign-ups come later.
This is actually your vault? Sign out and sign back in with the owner account.

Teach it once. Every assistant knows it.

Everything funnels through your approval.

Three steps, and the third one is the point.

1Connect your assistants

2They read one context

3You approve every change

Two surfaces: you, and how you work.

Who you are

What your assistants know

How your assistants do things

And it is always yours

Building this for a team?

You

Who you are

What your assistants know

Recently removed

Your lists

Ask

Projects

Inbox

Activity

Connections

Your connections

Prime the assistant to use it

Setup steps

Packs

People

Skills

Nothing matches

Add to your vault

This vault is private. Yours would work the same way.