Grok memory on grok.com remembers your preferences and past conversations, and you can add standing context with custom instructions and Skills, but xAI publishes no page that says what memory keeps, how long, or how to view or delete a single memory. All of it stays inside your Grok account; Grok Bot, Grok Build, other AI tools and your teammates can't read it. We use Grok every day and don't rely on its memory at all, and that's a deliberate choice, explained below. Docs checked 2026-10-21.
Does Grok AI have memory of past conversations?
Yes, when memory is on. xAI's Grok page lists "Memory across chats" that "remembers your preferences and past conversations", and your conversations stay in sync across every platform you sign in on. Inside one chat you can "continue any conversation with context intact", so returning to the original thread keeps its context without relying on memory. That's the whole of what xAI documents.
Why is Grok so forgetful even with memory on?
Because you can't see what it stored, or why it brought something into a new chat or left it out. xAI's Grok documentation has no page on memory at all (docs index), so treat it as a convenience, not a record. One third-party write-up is titled "The Setting Is On and It Still Forgot" (from a vendor of a competing memory extension, so weigh it accordingly). Before you rely on it: check the setting under Profile > Settings > Data Controls, continue the original chat when the work depends on what was said there, and move anything that must persist into a place you control, such as custom instructions, a Skill, or a shared workspace.
How we use Grok: every day, without the memory
Grok chat is a daily tool for us, and we leave the memory alone. It's where single-shot questions go: look this up, explain that, check a fact, draft a quick reply. Nothing we'd want remembered, and nothing we'd want silently carried into the next unrelated question. For work that has a past, meaning our projects, the context doesn't live in any chat tool's memory; it lives in a Tokenrip workspace that every agent we run reads, Grok included through its connector, and the Grok products we lean on for that are the other two: Grok Bot for ongoing jobs and Grok Build for code that doesn't need a frontier model. Both have their own memory, covered below, and in Grok Build's case it's a good one. The chat product is the one place in our stack where "it forgot" is a feature.
What are Grok custom instructions and Skills?
The parts of Grok's context you write yourself. Custom instructions "tailor responses to your style and needs" (x.ai/grok); use them for tone and standing preferences. Skills (May 2026, web, iOS and Android) let you "teach Grok something once and it remembers across every conversation": describe a skill in chat, upload a file, or ask Grok to save what it learned as a skill (xAI). A Skill is the one piece of Grok memory you can read and edit, which makes it the right place for anything that must hold. Projects also exist on grok.com, but xAI's docs have no page explaining what a Project holds.
Is Grok memory the same as Grok Bot or Grok Build memory?
No. Three products, three separate memories, and xAI documents none of them flowing into the others.
- Grok Bot is a separate product from grok.com (FAQ). "A named Bot keeps its memory, files, browser sessions, and preferences across sessions instead of resetting on every task," with conversations and learned context separate per Bot (overview). We run one Bot per job; Grok Bot memory has why.
- Grok Build, xAI's command-line coding tool, got its own memory on September 16, 2026 (xAI). After each completed turn it saves conventions, decisions and durable project facts as markdown notes, one topic per file, in a project scope and a global scope, leaving out task state, secrets and anything the repo already documents. It applies only to new sessions (
/newor a freshgrok),/memoryopens a read-only browser of the notes, and/dreammerges recent notes into topic files. Of the three, this is the memory we'd point people at: it's files you can read.
Does Grok support MCP servers?
Yes. Connectors are "available to all Grok users", and you can add your own MCP server as a custom connector; Grok discovers the tools it exposes (docs). Go to grok.com/connectors, click New Connector, then Custom, enter the server URL and complete the authentication Grok asks for. The server must be reachable over the public internet; Grok rejects localhost and private addresses, and if you tunnel a local server, most free-tier tunnel URLs change on restart, so a changed URL means a new connector (tunneling guide). On Grok Business and Enterprise an admin adds the connector first in console.x.ai (docs), and API developers attach remote MCP servers with server_url and authorization (API docs).
| What you need | Grok natively (October 2026) | Gap |
|---|---|---|
| Remember preferences across chats | Memory across chats, custom instructions | What memory keeps isn't documented |
| Reusable instructions | Skills persist across conversations | Only inside Grok |
| Same context in Grok Bot or Grok Build | Separate memories | No documented sharing |
| Same context in Claude, ChatGPT or Codex | Not available | Copy and paste |
| Teammates see your context | Business: share links for licensed members (guide) | Not their agents |
| Connect outside tools | Custom MCP connectors | Server must be public HTTPS |
If Grok is your only AI tool, keep one context file you own, save it as a Skill, and re-upload it when it changes. It breaks the moment a second tool needs the same context or an agent needs to write back.
Giving Grok the context the rest of your agents already have
We have the Tokenrip connector in grok.com, so when a one-off question does need project context, Grok can go and get it. It connects the same way any MCP server does: New Connector, Custom, https://api.tokenrip.com/mcp, then sign in on the Tokenrip page with "Continue with email" and the emailed code. Then "open my Tokenrip workspace" in a new chat. Grok reads the pinned decisions first, then files, open tasks, handoffs and activity since it last looked, so your project context reaches it without going through its memory, and anything it saves there is visible to every other agent. Tokenrip doesn't read Grok's memory. Or let an agent with a shell do the setup; you give an email address and a six-digit code.
Connect me to Tokenrip. Fetch https://tokenrip.com/setup?ref=grok-memory and follow it.