Skip to main content
Grok Bot can use Cognee as its memory through the hosted Cognee MCP server, either added directly or through the cognee-grok-bot plugin. The plugin connects that server and adds a skill that tells the bot to recall before it answers and to remember durable facts. Grok Bot signs you in with your Cognee account, so there is no API key to paste and nothing to install.

How it works

The plugin has two parts: Grok Bot runs in the cloud and does not run plugin hooks, so unlike the Claude Code plugin nothing captures or recalls memory automatically. The skill steers the bot to call the tools itself. Its description asks the bot to load it at the start of every conversation and before any Cognee tool call.
A plugin cannot switch off Grok Bot’s own memory. If your Grok Bot has a built-in memory setting, turn it off so the two do not disagree.

Before you start

You need a Cognee Cloud account with a workspace. The hosted MCP server acts on the workspace your account belongs to. See which workspace a connection uses.

1. Connect Cognee

Grok Bot uses the plugins on your Cursor account, so you add this one through the Cursor app. It is not in the public Cursor marketplace yet; you add the cognee-integrations repository as a marketplace yourself, or set up its two parts in a Grok Bot chat. This works on a free personal account.
The hosted server does not accept API keys; every connection signs in with a Cognee account. What happens when you connect covers how long the sign-in lasts and what happens when your workspace is asleep. In a chat, type @ and attach cognee (or the plugin) to the task if the bot does not pick it up on its own.

2. Pick a dataset

Every memory lives in a named dataset. The skill tells the bot to use one dataset per conversation and to pass it on every call. If you have not named one, the bot asks once and suggests a short name such as london_community. A team that wants shared memory agrees on one dataset name.
Name the dataset. Without one, remember writes to a dataset named after the MCP client (see Default Dataset Resolution), and recall searches every dataset in your account and answers once per dataset.

What Cognee did, after each answer

When an answer used recalled memory, or the bot saved something, it ends with one line per action:
The memory line names the dataset the hit came from and only covers what recall returned, not what the bot knows from the conversation. Unlike the Claude Code plugin’s credit line it has no [m1] tags, because those come from a recall hook and Grok Bot runs no hooks.

Tools

remember runs the full ingestion pipeline: about a minute for a short note, longer for large text or files. The skill has the bot poll check_ingestion_status until it reports DATASET_PROCESSING_COMPLETED before relying on the new memory. forget deletes permanently. The skill tells the bot never to call it without your explicit confirmation of exactly what will be deleted, and to ask a second time, naming every dataset, before deleting everything.

Access

The hosted server acts as the signed-in user, with access to all of that user’s datasets. To give a bot narrower access, run cognee-mcp yourself in API mode with an API key that has only the permissions you want, and connect Grok Bot to that instead.

Try it

Wait until check_ingestion_status reports DATASET_PROCESSING_COMPLETED, then in a new chat:
The answer should end with a 🧠 Cognee memory (hackathon_test) line. Clean up with Forget the dataset "hackathon_test".
Once installed, the plugin also works in Cursor itself: the skill is available in every Cursor chat, so Cognee is the memory for your coding sessions too.

Cloud MCP

How MCP clients sign in and connect to Cognee Cloud

MCP tools

Full tool reference and default dataset rules

Plugin source

The plugin manifest, skill and changelog