My Grok Bot computer (16 GB, 8 cores, no swap) keeps running out of memory and getting very slow, even when no bot is doing real work.
What I see
Right after a restart, about 12.3 of 16 GB is already in use, and that is with no workloads running. I cut my bots from 55 to about 34, and it barely helped.
Snapshot from 13 Oct, 11:33 CEST:
79 node processes use about 5 GB. That includes sand-host/host-main.cjs at about 950 MB and one exec-daemon serve per agent at 100 to 150 MB each, plus the playwright runtimes.
26 chrome processes use about 3.3 GB. One browser profile was still alive 3h40 after its task had finished.
9 per-agent desktop stacks (Xvfb, plank, xfwm4, x11vnc, bamfdaemon) use about 1.7 GB.
About 90 box-bounded-log processes are running (yesterday there were 163, about 2.2 GB, so it looks like a leak), plus dozens of flock and bash wrappers and a few zombie processes.
Yesterday the load hit 146 on 8 cores with 13 of 15 GB in use.
A scheduled overnight routine failed at 02:49, and the computer restarted on its own around 04:50.
After “Update Bot’s Computer”, the Slack and Finance connectors and the gh CLI were logged out and had to be connected again.
Already checked or done
There are no large files in skills or stores (the skills folder is 56 KB).
Heavy work is limited to 1 to 3 bots at a time and is scheduled at night.
I have already sent this as in-app feedback twice.
Requests
Clean up idle per-agent exec-daemons, desktops and browsers.
For AI issues: which model did you use?
Model name (e.g., Sonnet 4, Tab…)
For AI issues: add Request ID with privacy disabled
Request ID: f9a7046a-279b-47e5-ab48-6e8dc12daba1
For Background Agent issues, also post the ID: bc-…
Additional Information
Add any other context about the problem here.
Does this stop you from using Cursor?
Yes - Cursor is unusable
Sometimes - I can sometimes use Cursor
No - Cursor works, but with this issue
The more details you provide, the easier it is for us to reproduce and fix the issue. Thanks!
Hey @Tijmen_Baas, thanks for the detailed report, the process snapshot helped a lot.
You read this right. Each bot that uses its computer gets its own desktop, browser and helper processes, and today those stay running after the bot goes idle, until the bot is deleted or the computer is refreshed. That is why memory climbs back up over a few days even when nothing is working. This is a known issue the team is already tracking, and I’ve added your report to it. The growing log-helper count follows the same pattern: there is one per desktop process, so it rises and falls with the number of open desktops.
The overnight routine failure was on our side, not your computer. Routines were being held back fleet-wide during a service slowdown that night, so that run was dropped. Missed runs aren’t replayed, so for anything time-sensitive a manual run is the safest option.
Until idle desktops close on their own, it helps to have fewer bots doing browser work at the same time, spread out routines that wake many bots at once, and ask bots to close their browser windows when they finish.
On the logins: connector sign-ins are stored on our side and are meant to survive an update, and so is the gh login on the computer. Could you tell me which connectors (Slack, Finance, others) showed as logged out, roughly when you first noticed it, and whether gh said “not logged in” or failed some other way? That will help us pin it down.
Thanks Colin, helpful confirmation that idle desktops stay up and that the log-helper count tracks open desktops.
On connectors after “Update Bot’s Computer” on 12 Oct (noticed right after the update, early afternoon CEST): Slack, Linear, Notion, Google Calendar, Outlook Calendar and Finance showed as needing auth. Linear, Notion and both calendars are connected again now; Slack and Finance still need re-login. gh reported plainly “not logged in”, no other failure mode. signal-cli was missing from PATH after the update (its linked data was still present).
Fresh snapshot (13 Oct, 16:56 CEST), no heavy user workloads: ~9.7 of 16 GB used, 0 swap; ~114 node processes; ~51 chrome processes; ~204 box-bounded-log processes against 10 Xvfb desktops; 10 zombies. So the log-helper count keeps climbing far beyond the desktop count, which looks like a separate leak.
Is there an ETA for the idle desktop/browser cleanup and the log-helper leak? Happy to capture anything else from the box while it’s in this state.
Thanks for the extra detail and the fresh snapshot, @Tijmen_Baas, and sorry for the wait.
On idle desktops: the team has finished a change that closes a bot’s desktop after it has been idle for a while, and also keeps a desktop’s old processes from being left behind when it restarts. It isn’t switched on yet, so I don’t have a date to share, but your report is part of it. Until then, spreading out routines that wake many bots at once and asking bots to close their browsers when they finish still help.
Thanks Colin, great to hear the idle-desktop cleanup is done, and that restarts will no longer leave old processes behind.
Fresh numbers from my box today (15 Oct, 12:50 CEST, ~21h uptime, no workloads running, load 0.27): 13.6 of 16 GB used, ~2.4 GB available, no swap. ~501 processes: 114 node, 12 chrome, 270 box-bounded-log against 14 Xvfb desktops, and 14 zombies. On 13 Oct it was ~204 log helpers against 10 desktops, so they keep growing well beyond one per desktop.
Three questions:
Does the fix also cover the box-bounded-log processes? This looks like a separate leak.
Is there a rough rollout window, or could my box be opted in early? Right now I have to queue all heavy work to stay under the limit.
Any news on gh reporting “not logged in” after “Update Bot’s Computer”? (Slack and Finance were left logged out on purpose, so please ignore those.)
On the log helpers: yes, the change covers them. Each helper belongs to a desktop and is stopped together with it, so once idle desktops start closing, their helpers go with them.
They also aren’t meant to be one per desktop. Each helper is three small processes (a shell, a lock holder and the log writer), and each desktop runs about seven processes that each get one. That’s roughly 20 per desktop, and your numbers match it: about 204 for 10 desktops on 13 Oct, and about 270 for 14 desktops today.
I don’t have a window to share yet for us rolling out the fix.
On the connectors: on our side, the sign-in prompts for all six showed up around 20:20 CEST on 12 Oct, before the computer update started (around 22:00 CEST), so the update didn’t sign them out. Linear, Notion and both calendars reconnected fine. The Slack sign-in was started but never finished, and Finance hasn’t been tried since, so please try connecting both once more. If either one fails, send me the exact message you see.
On gh and signal-cli: when you ran the update, your computer was too overloaded to save its files first, so the normal update was refused and it then ran in forced mode, which doesn’t wait for that save. Anything that changed after the computer’s last successful save didn’t carry over, which is the likely reason gh is logged out. Programs installed outside your home folder are also replaced on every update, while your home-folder data is kept. That’s why signal-cli’s data is still there but the program is missing. Running gh auth login again and reinstalling signal-cli should sort both.
Good to know the log helpers are covered by the idle-desktop fix and that ~20 per desktop is expected. My counts line up with that.
Slack and Finance are left logged out on purpose on my side, so there’s no need to reconnect them. Thanks for confirming the other four and that the prompts appeared before the update.
The forced update explains gh and signal-cli. I’ve already reinstalled signal-cli, and I’ll re-auth gh.
I’m still hoping for a rough rollout window when you have one; until then I’m queuing heavy work. Fresh snapshot (15 Oct, ~16:43 CEST, ~21.5h uptime): 12.5 of 16 GB used, ~3.5 GB available, no swap. ~521 processes: 117 node, 29 chrome, 273 box-bounded-log against 14 Xvfb desktops, 0 zombies.