Does Grok Bot have a context limit? How Bot memory works, and the source-of-truth fix (unofficial guide)

The short answer

The official docs don't publish a context-window size for Grok Bot, so treat any exact number you see online as a guess.

What the docs do describe is memory: a Bot retains stable preferences, important facts, and summaries from its work, so it can keep a role over time without replaying every prior message. That means older details can come back as a summary, or not at all, rather than word for word.

So don't rely on memory for anything that changes or matters. Keep it in a source the Bot reopens, like a Google Sheet, a document, or a short decisions log, and ask it to check that source before important decisions.

This guide is independent and unofficial. It isn't affiliated with or endorsed by xAI, Grok, X, or Cursor. Features change often, so check the linked docs.

What does a Bot remember?

Per the docs, a Bot saves lasting things: how you like to work, rules you ask it to keep ("Always ask before you send an email"), facts about you such as your name and time zone, and a dated log of what it has worked on with you.

A Bot decides on its own when something is worth keeping and doesn't ask first. It often mentions a save in its reply, but not every save shows up in the chat. To make sure it keeps something, say so: "Remember that I'm out every Friday."

What a Bot saves in one chat, it remembers in all your chats with it. Each Bot has its own memory, but a few facts about you, like your name and time zone, are shared by all your Bots.

Logins, cards, addresses, and phone numbers belong in Vault, not in memory.

How do I see, correct, or remove a memory?

There's no memory screen for a personal Bot, so you ask in chat. To see it: "What do you remember about me?" To correct it: "That's out of date. I moved to Denver." To remove it: "Forget that I'm on the launch team."

If a fact shared by all your Bots is wrong, tell any of your Bots the right version. To remove a shared fact completely, ask the Bot that saved it.

Resetting a conversation starts the chat history over but keeps the Bot's memory, files, routines, and settings. Deleting a Bot deletes its memory, and deleting your account deletes all your Bots' memory.

A good monthly habit: ask each Bot what it remembers about you and its job, and correct anything stale. It takes two minutes and stops old assumptions from creeping into new work.

The fix: keep facts in a source of truth

The docs list four rules under "Memory is not a source of truth": keep changing facts in the source system, ask the Bot to cite or reopen current data for consequential decisions, correct stale assumptions directly, and back safety boundaries with approvals and Auto-review rules rather than memory alone.

In practice, give each kind of fact one home. Numbers go in the sheet or app they come from. Prices, policies, and client details go in a document. Decisions go in a short dated log the Bot appends to after each piece of work.

Then tell the Bot where those homes are and ask it to reopen them at the start of each task. Bots share one computer, so a file one Bot writes can be read by another, which also makes handoffs between Bots more reliable than relying on each one's memory.

Put the instructions themselves in a skill. A skill is a saved recipe the Bot reads every time it runs the task, so the method doesn't depend on what it happens to remember.

Long chats and big jobs

If one chat has grown huge and the Bot keeps mixing up old and new details, ask it to write a short summary of the current state and open questions into a file, then reset the conversation. Memory, files, and routines stay; the clutter goes.

Split big jobs into smaller tasks with a clear finish line, and have each one end by updating the decisions log. Each step is easier to check, and the next step starts from a written record instead of a long history.

For anything consequential, like money, contracts, or what you've promised a customer, ask the Bot to show where the fact came from. If it can't point to a source, treat it as a guess.