Grok Bot computer: RAM stays near full with many bots, and about 8 GB isn't attributable to any process

Support Bug Reports
, ,
1

Describe the Bug

Same symptom as Memory Bleeding. My Grok Bot computer has 16 GB of RAM and no swap. Available memory has dropped to about 2.2 GB several times, and on 12 Oct 2026 about 13 bots running browser sessions at once crashed it. I don’t have timings or exact symptoms for that crash.

What was measured:

  • 15 Oct, about 7:52pm SGT: total process memory was about 16.2 GB on the 16 GB computer, with about 1.2 GB free. It was split roughly like this:
    • about 30 always-open bot desktops, about 3.7 GB
    • about 195 log-watcher helper processes, about 3.5 GB
    • about 46 leftover Playwright helper processes, about 3.4 GB
    • about 23 bot workspaces, about 2.4 GB
    • the main app, about 0.7 GB
    • actual Chrome, only about 1.6 GB
    • These persisted for days, including through Update Bot’s Computer.
  • After I deleted some bots (15 Oct evening): 13.7 GB used and about 2.3 GB available. Desktops dropped from 31 to 28.
  • 23 and 24 Oct: about 13 GB used and about 2.2 GB available, with 4 to 5 desktops. Visible processes added up to only about 5 GB, so about 8 GB of used memory was not attributable to any process I could see.
  • 28 Oct, about 5:17pm SGT: MemTotal 16.4 GB, MemFree 5.7 GB, MemAvailable 7.9 GB, Cached 2.7 GB, AnonPages 3.4 GB. Available memory was much better than on 23 and 24 Oct.
  • Kernel messages: the virtio_balloon module is loaded, and dmesg showed repeated virtio_balloon: Out of puff! Can't get 1 pages messages on 23 Oct. There are no OOM-killer lines, but the logging services were killed repeatedly (systemd-journald ... status=6/ABRT).

What I tried:

  • Deleting bots, staggering routines and pausing bots before Update Bot’s Computer. None of it fixed the problem. Each bot costs roughly 300 to 600 MB while it exists, even idle, and I couldn’t safely close the app’s own helper processes.
  • I sent product feedback on 23 Oct.

Request:

  1. Please explain what the roughly 8 GB of unattributed memory is. I suspect the host VM’s balloon, but I haven’t confirmed that.
  2. Please fix the leftover desktops, log watchers and Playwright helpers that stay running after a task or after Update Bot’s Computer.

Steps to Reproduce

.

Expected Behavior

.

Operating System

Windows 10/11

Version Information

Version on 12 Oct till now, on Android phone and desktop app

Does this stop you from using Cursor

No - Cursor works, but with this issue

7

Your guess on the 8 GB is right. It’s held by the host’s memory balloon, not by any process on your computer. When the computer wakes from sleep, the host can take back up to about 10 GB of the 16 GB and hands it back in steps as the computer needs memory. Linux counts that memory as used even though no process owns it, and the Out of puff lines in dmesg most likely come from that driver. So it isn’t a leak, though right after a wake it can take a little while to come back.

The leftover desktops and log helpers are a known issue the team is aware of, and I’ve added your report to it. Since about 22 Oct, a bot’s desktop and its log helpers close automatically after around 30 minutes idle, which likely explains the drop you saw on 23 Oct.

If Playwright helpers still pile up after tasks finish, could you post a count and how long after the task you checked?