A Mac that ran out of room 436 times

This is the Mac RefreshYourMac is built on: an M1 with 16 GB of memory and a 245 GB drive. Since September it has also been where I make things with AI: a narrated Sherlock Holmes series for YouTube and a three-part audiobook, both produced with Codex, alongside the day job of building apps with Claude Code. RefreshYourMac keeps a record of every clean, so the cost can be read rather than guessed. On 3 November the drive had 15 GB free.

Running out of room is the problem I built RefreshYourMac for: a way to find the data I could safely clean, and to know what each clean-up would cost before anything was deleted. These are the numbers from my own Mac since.

436 times in six weeks

RefreshYourMac can clean the Safe categories on its own when free space falls below a line, here 10 GB. From 23 September to 3 November it did so 436 times, up to 24 times in a day.

386 GBfreed across all 535 cleans, 15 September to 3 November

That is not 386 GB of different files: most of it is caches that refilled and were cleared again, which is the point. Working with AI generates data continuously, and on a 245 GB drive it is the difference between room to work and none. Here is where it came from.

A YouTube series: close to 7 GB of renders, in four versions each

Codex rendered every episode of the Sherlock Holmes series to 1080p, and every change meant a new render. By October each episode existed in up to four versions:

• An untrimmed cut, and for some a title-untrimmed one.
• A 1080p render, and then a "measured" 1080p final.
• A second set for A Scandal in Bohemia, The Red-Headed League and both parts of The Hound of the Baskervilles, re-recorded in a different synthetic voice (Neural2-D).

On 11 October RefreshYourMac's Large Files list moved them to the Trash in three passes: 4.45 GB, 1.4 GB (one installer download among the episodes) and 993.5 MB. Emptying the Trash that evening freed 6.91 GB. A month earlier, on 15 September, emptying it had freed 723 MB, including an archive of every episode's narration and a screen recording.

None of those files were the uploads themselves. They were the versions that led to them, kept because nothing in the process ever deletes the last render when it makes the next one.

An audiobook: 748 MB of working audio behind 193 MB of finished episodes

The audiobook, three episodes so far (The Seventeenth Note, The Drowned Man's Watch and The Knight in the Locked Chapel), was voiced one line at a time, as a WAV file per line of the script, then joined and compressed. The second and third episodes were recorded twice. The finished episodes come to 193 MBof .m4a audio; the per-line takes behind them came to 748 MB, about four times as much, all still on the drive two months later. They went to the Trash on 3 November, after checking each finished episode was where it should be. The scripts that made them stay, so any line can be voiced again.

The tools themselves: caches that come back

• Codex runtimes, 13.4 GB over eight cleans. Codex downloads the runtimes it uses to run code into ~/.cache/codex-runtimes, and downloads them again after they are cleared, about 1.6 GB each time.
• Codex updates, 8.8 GB over four cleans. The updates Codex downloads collect in its cache and build up again after a clean.
• Claude Code's history, 5.4 GB and growing. ~/.claude was 1.73 GB on 1 October and 5.4 GB on 3 November; 5.3 GB of it is session transcripts. It is a record rather than a cache, so it only grows.
• Codex's history, 3.0 GB in ~/.codex, up from 2.94 GB on 1 October.

macOS's own AI: 12.7 GB of half-built models in one day

The largest single item was not from any tool I installed. On 1 November RefreshYourMac found12.7 GB of half-finished copies of the AI models macOS's Photos analysis compiles for each Mac: 186 of them, around 60 MB each, made in a single day by a service that kept quitting before a compile finished and starting the next one from scratch. Each abandoned copy stays where it was left. The same day, the app cleared 626 MB of the AI models Lightroom and Photoshop download for Denoise and masking, and 208 MB of Chrome's on-device AI models.

Apple Intelligence also takes a share of the drive that no app shows. Apple's own figure is up to 8 GB on most supported Macs, including this M1, and up to 14 GB on a Mac with an M3 or later and at least 12 GB of memory, and it notes the space used can be more than that for a while during and after a software update (Apple Support).

What I would do differently

• Delete the working renders once the upload is done. Keep the final file and the script or project that made it; every intermediate version can be made again.
• Keep voice takes only until the episode is mixed. Per-line WAV files are the largest thing an audio pipeline makes and the least useful afterwards.
• Look at the Trash and the Large Files list after each project, not when the disk is full. Seven gigabytes of episodes sat there unnoticed for weeks.
• Leave the regrowing caches to a schedule. Codex runtimes and updates come back by themselves; clearing them by hand every week is work a scheduled clean does better.
• Decide how much history to keep. ~/.claude and ~/.codex are records of every session. Copy them elsewhere before trimming them, and move them to the Trash rather than deleting outright.

Where the app comes in

Every number above came from RefreshYourMac's own history. Large Files and the Downloads folder are rated Review before deleting and go to the Trash; Codex runtimes and updates, and unfinished model compiles, are Safe for auto removal; Claude Code and Codex history are kept for you to decide. Where AI tools keep their data, and what is safe to delete has the folders and the commands to size them yourself.