[Windows Desktop] Memory grows sharply with multiple windows and long-history threads in 26.810.41047

Summary

On Windows, Codex Desktop's memory usage increases sharply when multiple app windows are open and a long-history/tool-heavy thread is loaded. The memory footprint remains elevated after the thread/window has settled instead of returning to the empty-task baseline.

This report is based on a local A/B measurement and is not a claim that this build is the sole cause.

Environment

• Codex Desktop About: 26.810.41047 (released 2026-08-14)
• Installed Windows AppX package: OpenAI.Codex_26.810.4967.0_x64
• ChatGPT.exe file version: 151.0.7922.137
• Windows x64 (exact OS build not captured)
• The machine remained usable during the reproduction; this was elevated app memory, not a system OOM.

Reproduction / measurements

1. Restarted the PC and Codex, then opened an empty task.
   • Task Manager showed the ChatGPT group at approximately 777.6 MB.
2. Opened a second Codex window and loaded a long-history/tool-heavy thread.
   • ChatGPT group/full working set reached approximately 1.5–1.6 GB.
3. Opened a third Codex window.
   • Task Manager screenshot showed 2,009.9 MB and 13 processes.
   • A live 57-second sample measured full working set 2.019 GB initially, 1.926 GB at the end, and 1.89 GB shortly afterward.
   • Private working set during that sample was 1.27–1.38 GB.
   • The post-settle process count was 11.
4. During the third-window sample, available RAM stayed around 16.5–16.9 GB, pagefile usage was about 1.36%, CPU stayed roughly 1.7–14.7%, and disk queue stayed near zero after the initial open burst.
5. In an earlier 20-second post-restart sample, no git.exe processes were observed, so this reproduction did not depend on a simultaneous Git polling storm.

Expected behavior

Opening another window or loading a long thread may temporarily increase memory, but usage should stabilize and be released when the window/thread is no longer active.

Actual behavior

Each additional window/history load adds a substantial amount of memory. With three windows, the ChatGPT group remained near 1.9–2.0 GB after settling, well above the empty-task baseline.

Questions

• Is this expected renderer/history retention behavior?
• Are long thread histories virtualized or unloaded when inactive?
• Are there diagnostics or a recommended workaround for multi-window use on Windows?
• Please let me know if a sanitized process/memory trace would help.

Start by reproducing the reported Windows multi-window and long-history scenario using the measurements in this report, then collect a sanitized process or memory trace if available. Done means identifying the source of retained memory and confirming that usage stabilizes or is released after windows or threads become inactive.