
What happened
Running Claude Code sessions via TerminalHub on one 32GB Windows PC, the user logged 33 heavy jobs over three days and summarized 30; all finished, including 9 that dropped free memory under 1GB, and the disk-saturation watcher never fired.
Why it matters
The old rule (stop when free memory falls below 1GB) would have halted 9 of those 30 jobs unnecessarily, since low free memory did not by itself predict the SSD stalling as the author had assumed.
Summaries like this, every morning.
The author had previously managed heavy jobs by reserving memory before starting them and stopping when free memory fell below 1GB. That approach assumed low free memory was a proxy for the two problems that actually matter: the PC becoming sluggish due to disk paging, and processes crashing when the commit limit (the total memory plus page file Windows will guarantee, 47.5GB on this machine) is reached. The three-day log shows free memory was a poor proxy for the first problem: a 40-minute ComfyUI generation that bottomed out at 11MB free memory averaged only 14% disk usage and raised page file usage by just 3.8 points. The distinction between the two problems appears in the one job that stopped, a roughly 57-minute ComfyUI run with no other reservations: disk usage averaged 30% and never met the saturation condition, but commit reached 46.1GB and the script's own line halted generation. The author also found the watcher's cadence (one 60-second window every 10 seconds) could miss short stalls, and fixed it to examine every window. The thresholds (90%, 60 seconds, +0.3 points) were derived from a single 24-image test, and the author notes they would likely differ on other SSDs or page file sizes. The rules were built for this specific PC with an integrated GPU sharing the same 32GB, not for a machine with dedicated GPU memory, and the watcher relies on Windows typeperf counters.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
AIworker analyzed recorded consultations and found ChatGPT's share of new corporate AI adoption inquiries fell…

The piece sorts agent-driven video into video-generation models like Seedance, front-end code such as Three.js…

Levent Alpöge published a 4-page remark.pdf saying he believes he found a Riemann Hypothesis proof with Claude…

On 2026年11月1日, EveryInfra tested Claude Opus 5.5, Sonnet 5.5, and Fable 5.1 and found forced tool calls return…

A Zenn explainer lays out ReAct, the 2022 Google Research pattern in which an LLM alternates Reasoning and Act…

A developer installed a reviewer subagent that writes a sha256-tagged PASS/FAIL verdict to .review/verdict.jso…
