What is the best way to go about exporting/backing up each bot’s individual memory, as well as the shared bot memory?
I’m not referring to backing up chats, I’m not interested in those.
I’m looking to export all of the persistent data, or as Dean Rie put it on another thread:
The bot’s durable state (profile, memory, routines, skills, connectors, files on the shared computer)
I don’t have any custom skills whatsoever currently, AFAIK.
I thought about having each bot create a skill.md file for each topic in its individual memory, and having the Chief of Staff create a skill.md file for each topic in the shared memory.
I saw that it is possible to transfer a certain topic from one bot’s individual memory to another bot’s individual memory. I have done this when I created a new bot for a niche purpose. I had another bot doing that niche work before. When I made the new bot, I watched as it “transferred” about 55 memories on that niche topic to the new bot; erasing them from its own individual memory as it went along.
So it looks like it is certainly possible to ascertain which memories relate to each topic, so I guess it is possible to make skill.md files for each topic?
I cannot say enough how much I absolutely love Grok Bot! It is hands-down the best piece of tech that I’ve ever used! (Using Lauren’s pstack). However the pricing currently makes it impossible to use (1 week’s usage lasts me 8 hours).
I’m considering trying/switching to Hermes (despite its flaws that I’ve heard about), so I’m going to do this export from Grok Bot on my last week (currently 2 weeks in to a 1 month Cursor Ultra subscription).
There’s no one-click export yet, but you can back everything up by just asking the bots - a bot’s memory lives as plain text files on its own computer that it can zip up and hand back.
In a chat with one bot, paste: “Please make a full backup of your durable state. Make a folder backup/<your name>, copy your complete memory (profile + every memory-log entry, including older months) into memory.md, list your routines (schedule + prompt) in routines.md and your connectors in connectors.md, then zip the folder and send it here as an attachment.” Download the zip when it appears, and repeat for each bot.
For shared memory + skills, ask your Chief of Staff: “Copy the shared memory all bots can see into shared-memory.md, zip the skills library folder, and send both here as attachments.”
For other files on the shared computer, tell any bot which folders matter and ask it to zip and send them (attachments cap at 25 MB - for bigger folders ask it to split the zip).
Two notes: connector logins can’t be exported (the list is just a record), and your SKILL.md-per-topic idea works too - add “also write a SKILL.md for each distinct topic in your memory and include it in the zip” to step 1.
A proper export/backup is something we’d like to offer - I’ve noted your request (and the pricing feedback) for the team.
Thank you so much for your detailed reply! I will try this soon.
I’ve seen that you guys have been trying to fix the burn rate of the weekly usage, over the weekend - it seems to be considerably better!! I got a reset on the 5th and I’ve used 85% in 2 days, compared to 100% in 1 day. I still haven’t got it running completely autonomously (while I sleep) though, otherwise it would still burn 100% in 1 day.
From my ChatGPT routine that’s monitoring the high burn rate of weekly usage:
Meaningful update — 5 September 2026: there are two major developments.
1. Grok Bot received another full usage reset for all users. The official @bot account announced: “We’ve reset usage limits for all Grok Bot users. Enjoy!” Elon Musk then quote-posted it at 20:19 UTC / 21:19 BST on 5 September, saying “Grok @Bot usage limits reset.” This is another off-cycle reset after the resets on 26 August and 1 September.
2. More importantly for the rapid-depletion problem, the Grok Bot/Cursor team says it has now shipped substantial token-efficiency improvements. Jediah Katz, who describes himself as building the agent at Cursor and @bot, says the team spent the previous two weeks improving routing, caching, dynamic context usage and overall token efficiency. He says Grok Bot usage should now go about 10% further on average, and up to 35% further for certain heavy users.
Matt Palmer, who works with Cursor/Grok Bot, summarizes the change even more explicitly: “The @bot harness is now 10-35% more efficient and usage has been reset for the long weekend.” Lauren Tan also says Grok Bot usage “should go further now” and that the team will continue improving efficiency.
This appears to be the first quantified deployed improvement directly addressing the unusually fast Grok Bot usage depletion we’ve been tracking.
It linked a few different Twitter posts, but they haven’t pasted.
Just to follow up on this, since the team improved the usage, I’m seeing it go twice as far. I probably won’t be trying Hermes. Grok Bot is super awesome!
I’m still going to try to build the skills though. Since my original post, I’ve started getting the bots to maintain a directory of skills Markdown files in a Git repo And one of my skills is /update which grabs the latest skills from the Git repo and installs them in Grok Bot
I haven’t tried to export memory yet, but I do think it would great to codify a lot of my memories in to skills. It will no doubt be a fairly involved task to do, but I might try it eventually lol.
I’d keep the backup and the skill library separate. Skills are reusable behavior; memory is user or task state with provenance, timestamps, scope, and often an expiry, so converting it wholesale into SKILL.md can freeze stale facts into instructions. Export the raw memory plus a manifest, test a clean restore, and promote only stable, reviewed procedures into skills.
@DannyB love hearing that, and glad the efficiency work is landing for you. Your Git repo of skills with an /update skill to pull the latest into Grok Bot is a great pattern.
On the memory side, @izgorodin’s point is a good one: keep the raw memory backup separate from your skills, and promote only stable, reviewed procedures into SKILL.md. Memory is timestamped, scoped state that can go stale, so converting it wholesale into instructions can freeze old facts in. Export the raw memory as its own zip first, then cherry-pick what’s worth codifying.
I got a reset on the 5th and I’ve used 85% in 2 days, compared to 100% in 1 day. I still haven’t got it running completely autonomously (while I sleep) though, otherwise it would still burn 100% in 1 day.
And one of my skills is 