Local AI Windows 11 Disk Cleanup Misses Recycle Bin, Crash Dumps and NVIDIA Cache

The experiment makes a useful distinction for anyone considering AI-assisted disk cleanup: locating a large file, understanding its purpose, and recommending its removal are three different jobs. A model can appear competent at the first while remaining unreliable at the other two.

Crucially, this assistant had no delete button. That was the strongest part of the design.

A drive inventory, not control of the drive​

Raj’s test began with a familiar problem: a 1TB SSD repeatedly approaching capacity, followed by a manual review in WizTree. Instead of asking the model to scan Windows directly, he exported the file inventory into a SQLite database.

A Python script connected that database to a local model through llama.cpp. Four read-only tools could retrieve:

• The largest files.
• Folder sizes.
• Files within a selected folder.
• Storage totals grouped by file type.

Each lookup returned at most 60 rows. The model could not execute code, browse the web, inspect file contents, or delete anything. Downloads and some personal application data were excluded, so this was a privacy-filtered inventory—not an exhaustive audit of every file.

XDA identifies the model as Qwen3.6-35B-A3B, using UD-IQ4_XS quantization and a 32,768-token context window. Raj describes it as an Apache 2.0-licensed mixture-of-experts model with 35 billion total parameters and 3 billion active parameters. It ran on an RTX 4070 Ti Super with 16GB of GDDR6X memory, with some expert layers offloaded to system RAM.

Reported sampling settings were temperature 1.0, top-k 20, and top-p 0.95. Thinking mode was enabled, but earlier reasoning was not returned to the model, and older lookup results dropped out of its visible conversation.

The instructions required paths, sizes, and explanations, prohibited recommendations involving personal files, and told the model to acknowledge uncertainty. Those are sensible boundaries. They did not guarantee sensible conclusions.

The biggest recommendation offered temporary relief​

The final report’s claimed savings were dominated by a 16GB Microsoft Flight Simulator rolling cache.

Raj’s objection was practical: in his setup, launching the simulator would recreate it. Removing a regenerating cache may reclaim space immediately, but that is not the same as solving a recurring capacity problem.

The model also recommended approximately 496MB of Microsoft Edge caches and 150MB of temporary installer leftovers. Those were more useful candidates in Raj’s assessment, although his description of them as unconditionally safe deserves qualification.

Microsoft’s Disk Cleanup documentation distinguishes temporary setup files left by a setup program that is no longer running and says temporary files unmodified for at least a week can be safely deleted. That is narrower guidance than treating everything in a Temp directory as disposable. A folder name is a clue, not a deletion policy.

Another recommendation targeted C:\Program Files\llamacpp, approximately 1.1GB. The model suggested removing the AI toolkit if local models were not being used—while Raj was running a local model through llama.cpp.

There was a complication: the directory was reportedly a second installation. It might have been redundant, but the inventory alone did not establish that. The better recommendation would have been to verify which installation the script depended on before removing the other.

The model also attributed substantial game storage to Epic Games, despite Raj saying his three largest installed games—Red Dead Redemption 2, Star Wars Jedi: Survivor, and Resident Evil Requiem—had not come from Epic. The report does not establish why that attribution appeared.

It discovered the larger opportunities—and omitted them​

The lookup logs told a more interesting story than the final answer.

The NVIDIA figure differs slightly between XDA’s narrative and table, hence the range. Likewise, “about 17GB” is the author’s summary, not a precisely reconciled total or a measurement of space actually reclaimed.

The omission of the Recycle Bin was particularly striking: the model inspected it on its second turn. Microsoft confirms that files there are not permanently removed until the bin is emptied. Reviewing its contents therefore matters even when they have already been marked for deletion.

The crash report required a different judgment. Microsoft documents C:\Windows\LiveKernelReports as a typical location for live dumps and explains that these files support debugging with WinDbg. A large dump may be a cleanup candidate after an investigation, but its size does not make its diagnostic value disappear.

The failure was incomplete reporting, not simply incomplete discovery.

Was context management the culprit?​

Raj suggests that earlier discoveries may have fallen out of the model’s visible context before it generated its answer. That explanation fits the script’s reported behavior, but it is not a proven diagnosis.

A larger context window was not tested as a comparison. Nor was the published result a benchmark: earlier attempts encountered script bugs, context overflow, and empty answers, and only one run with the finished script and final database was reported.

Raj recorded generation speeds of roughly 65–70 tokens per second. His concern that a larger context could increase memory pressure and slow inference remains a proposed trade-off, not a measured result.

An engineering improvement worth considering would be a persistent candidate ledger maintained outside the conversation. Each discovery could retain its path, size, supporting lookup, uncertainty, and inclusion status. That would make omissions auditable rather than relying on the model to remember everything. It is a design recommendation—not a demonstrated fix for this experiment.

A safer Windows cleanup workflow​

Windows users need not reproduce the AI setup to act on the useful lesson.

Microsoft documents Settings > System > Storage > Storage Sense for configuring temporary-file cleanup, Recycle Bin retention, Downloads cleanup, cloud-content handling, and schedules. Review those settings deliberately: Downloads deletion and making cloud-backed files online-only have different consequences.

For Edge, use its browsing-data controls and select Cached images and files, rather than deleting a profile directory. Microsoft distinguishes cached content from cookies, passwords, and other browser data; clearing cache need not become an accidental account-and-preferences purge.

A cautious approach is to:

1. Review large candidates and identify the owning application.
2. Confirm that Recycle Bin contents are genuinely unwanted.
3. Preserve dumps needed for an active investigation.
4. Prefer Windows or application cleanup controls.
5. Recheck free space after cleanup and again after normal application use.

That last check distinguishes temporary breathing room from durable savings.

An assistant, not a disposal authority​

Raj reports no system-damaging recommendation in this run, but the model never executed a deletion. The experiment therefore demonstrates a useful read-only safety boundary—not proven autonomous cleanup safety.

The balanced conclusion is that local AI may help generate an inspection shortlist, while a human still validates dependencies, retention needs, and likely regrowth. For Windows administration, Microsoft also provides Storage Sense configuration through Intune, Group Policy, and CSP, offering explicit policies rather than conversational guesses.

Let the model point at the clutter. Keep the final decision—and the delete button—somewhere more accountable.