Grok Bot Memory + Which AI Models Hold Context: What Real GoldieBench Scores Say About Recall

Grok Bot memory and a model's context window are two different things, and GoldieBench only measures one of them.

Memory is what a Bot saves and brings back across your chats.

A context window is how much text one model can hold in a single request.

This page reads the live GoldieBench board on 2 November 2026 and says which scores help you here and which don't.

Want help choosing a model stack? Book a free SEO strategy session with my team.

Grok Bot memory is saved facts, not a bigger context window

Per the Grok Bot docs, a Bot keeps preferences, rules, facts like your name and time zone, and a dated log of its work.

Since v0.75.2, a save made in one chat carries into every other chat you have with that Bot.

It keeps summaries rather than replaying every prior message.

That's a storage feature of the app, so it works the same whichever model answers.

You can't choose the model behind your Bot

The Grok Bot FAQ says there's no model picker.

It picks a backend model per task, first-party or third-party, and the choice can change from one request to the next.

It also doesn't show which model handled a request.

So no score below changes what your Bot runs on.

The scores help in two places: setting expectations for xAI's models, and choosing models for the agents you run beside Grok Bot.

The xAI rows on the board

These rows are quoted from the GoldieBench leaderboard and the model pages, read on 2 November 2026.

Grok's top task on its model page is Twilightvale, a game, at 9.5 with gold.

The Grok 4.7 page also lists outside reference figures, such as a 1,695 Elo on GDPval sourced to x.ai, kept apart from the medal scores.

Context windows on the board, next to the leaders

MiniMax M3 is the one leaderboard row that lists a context window as its reference figure.

A long window helps when one request has to read a big document at once.

It doesn't give a Bot memory between sessions.

Where saved memory meets the model

The inference hooks docs, an Enterprise beta, show how memory reaches a model.

Context Grok Bot gives the model, such as rules, shared memory and connected apps, usually arrives in an attachment block's text.

So whichever model answers, your saved facts ride along as part of the request.

The same page says ancillary calls, such as saving memories, aren't sent to the hooks.

That's why a model with a long window can read your memory, but doesn't create it.

What a 0–10 one-shot score tells you about memory

Very little, directly.

The methodology page says every model gets the same fixed prompt, one shot, and must return a single self-contained HTML file.

That tests building in one pass, not recalling a rule you set last week.

What it does tell you is how well a model follows a dense brief first time, which matters when a Bot works from saved rules.

Pairing models with a memory-heavy job outside Grok Bot

If you run your own agent next to Grok Bot, you do choose the model.

• Careful rule-following: Claude Opus 5 averages 8.27 across all 50 tasks, every one scored.
• Blended answers: Hermes MoA averages 8.17 over 47 scored tasks.
• Huge single reads: MiniMax M3 shows a 1,048,576-token context window with a 7.97 average.
• Live X context: the Grok model page says it's used for real-time content workflows that need the X timeline.

Keep the same split Grok Bot uses: saved rules in a memory file, live facts fetched fresh.

A high score with no memory file still forgets you between sessions.

A modest score with a tidy memory file and a fresh data fetch often serves a business better.

Voice and group chats change what the model sees

Calls aren't a blank slate either: the docs say a Bot on a voice call can lean on past calls plus your latest chat.

Group chats let your Bot use things you said privately, while each Bot keeps its own memory.

Both are app features, so they work the same whichever model the Bot picks.

If you build your own agent, you have to design that carry-over yourself, which is where a model with strong rule-following earns its score.

Test recall yourself, since no board measures it

GoldieBench has 50 tasks and 33 models on the board, and none of them is a memory test.

So run your own three-step check on any Bot.

1. Tell it one lasting rule and ask it to remember it.
2. Open a new chat with the same Bot and ask "What do you remember about me?"
3. Correct one item, ask it to forget another, and ask again.

The Bots docs say there's no memory screen, so this chat check is the audit.

Keep changing facts out of memory

The docs say memory isn't a substitute for an authoritative source.

Ask the Bot to reopen current data for big decisions, and back safety rules with approvals, not memory alone.

Logins, cards and contact info go in Vault.

On teams, the changelog says every linked teammate gets their own memory and connectors on a Team Bot; my Grok Bot teams breakdown covers the model side of shared Bots.

Why the board still matters for Bot owners

Grok Bot users can't route requests, but most also run other agents.

The board tells you which model to put on the jobs Grok Bot doesn't cover.

Fusion leads at 8.59 across 47 scored tasks, and Claude Opus 5 sits second at 8.27.

Grok, at 8.09, sits within half a point of the leader, which is a fair sign of how capable xAI's own model is.

Reading the board without fooling yourself

Use the averages to pick models for one-shot builds.

Use context windows to judge single long reads.

Use your own recall test for memory.

And remember a Grok Bot reply may come from a model that isn't on this board at all.

Inside the AI Profit Boardroom we compare these models on real business jobs every week.

Also On Our Network

• Agent Operatorsthe five-step setup for seeding, importing and cleaning a Bot's memory
• AI Income Deskwhat Bot memory means for a small business and its staff
• AI Tool Verdictthe eight-criteria scorecard rating Bot memory out of 10
• agentos.guidefive Agent OS prompts that write, audit and tidy Bot memory
• aiprofitboardroom.comthe complete Grok Bot Memory guide on the AI Profit Boardroom blog

Run the three-step recall test on your main Bot today and book a free strategy session if you want help pairing models with Grok Bot memory.