What “junk files” are, and what Windows already deletes

Subject What a cleaner counts as recoverable, and what the operating system removes without being asked Covers cleanmgr · Storage Sense · WinSxS · hard links · prefetch · the page file We ran it No. We hold no copy and distribute none. Sourcing Microsoft documentation · a Microsoft servicing blog · one archived vendor page Ledger rows D-01 … D-09, plus D-10 on reserved storage Open questions 4 · listed Published 2026-08-06 Last verified 2026-08-06

One storage extent drawn as a single horizontal bar. Left of the dashed vertical boundary the extent is a hollow outline: the size a directory listing reports, which counts every hard-linked file once per name. Right of the boundary the same extent is drawn solid, sized to the bytes actually occupied on the volume, with hollow blocks punched through it for the part a supported cleanup can reclaim. The boundary marks **reported ↔ occupied**, and the gap across it is the arithmetic this page is about.

A scan that reports 4.7 GB of recoverable junk has published a number and withheld both of the things that would make it checkable: the definition of junk it used, and the method by which it measured. Neither is a nicety. Junk is not a Windows concept — the word appears nowhere in Microsoft’s reference for its own disk-cleaning tool, which instead enumerates ten categories and defines each in a sentence. Every product using the word is shipping a private list, and the size of the number it produces is a property of that list rather than of your machine.

Verdict SUPERSEDED for scan-and-clean junk removal (D-01) · NO-OP for “system trace files” (D-08) In one sentence Windows deletes most of this list on its own schedule and for nothing, the largest headline figure in the category is an artefact of measuring hard links twice, and the honest name for what is left is disk space, which is a different quantity from speed.

§1“Junk” is not a Windows concept, so every recovered-space figure is a private measurement

Uses of “junk” in the cleanmgr reference: 0
Categories named there instead: 10, each defined in one sentence
Published method behind a scan total: NOT ESTABLISHED — see §8

Microsoft’s command reference for cleanmgr opens by saying the tool “clears unnecessary files from your computer’s hard disk”, and then does the work that makes that sentence mean something: it lists the categories a cleanup can select — Temporary Setup Files, Downloaded Program Files, Temporary Internet Files, Old Chkdsk Files, Recycle Bin, Temporary Files, Temporary Offline Files, Offline Files, Compress Old Files, and Catalog Files for the Content Indexer — and defines each. Its guidance on the largest of them is a rule you can apply yourself: “You can safely delete temporary files that haven’t been modified within the last week.” That is a definition and a threshold. It is falsifiable.

Claim Scan and clean up computer junk files by Windows Winset. Archived Windows Winset capability index, tlwinset.com/help.htm, capture 2012-08-24. Checked 2026-08-06 · Ledger row D-01

Microsoft is explicit that a number of this shape is a classification trigger. Under Unwanted software → Lack of choice, the current criteria page — ms.date 2026-01-29 — names three behaviours: software may “display exaggerated claims about your device’s health”, “make misleading or inaccurate claims about files, registry entries, or other items on your device”, and “display claims in an alarming manner about your device’s health and require payment or certain actions in exchange for fixing the purported problems.” That is the operating system’s vendor describing the scan, not the cleaning.

What we could not establish. We looked for any product in this category that publishes its junk definition and its measurement method, and we did not find one; we looked at Microsoft’s own documentation set and at the single archived vendor page this domain carries. That absence is a finding rather than a conclusion, because we cannot search every product page that has ever existed, and we say so here rather than in a footer.

§2Windows deletes most of this list on its own schedule, and the scriptable version has shipped since Windows 2000

Storage Sense default cadence: 0 — during low free disk space
Recycle Bin default threshold: 30 days
Downloads default threshold: 0 days — meaning never

Two first-party mechanisms cover this ground and both are free. Disk Cleanup is scriptable: cleanmgr /sageset:n writes a profile as StateFlagsNNNN values under HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows\CurrentVersion\Explorer\VolumeCaches, and cleanmgr /sagerun:n executes that profile against every drive. Each selected box is a REG_DWORD of 00000002 and each cleared one is 00000000. One correction we owe the record, because we expected to write the opposite: Disk Cleanup is not deprecated. We read Microsoft’s deprecated-features list for Windows client on 2026-08-06 and neither Disk Cleanup nor cleanmgr appears on it, in any row, at any announcement date. The widely repeated 2018 deprecation is not on a Microsoft page we could load.

Storage Sense is the newer of the two and its defaults are published as policy, so they can be quoted rather than guessed. Microsoft’s Storage policy CSP states that “by default, Storage Sense is automatically turned on when the machine runs into low disk space and is set to run whenever the machine runs into storage pressure”, and that temporary-file cleanup is on unless someone turns it off: “by default, Storage Sense will delete the user’s temporary files.” The Recycle Bin threshold defaults to 30 days. The Downloads threshold defaults to zero, which the same page glosses as “never deleting files in the Downloads folder”, and the cloud-content dehydration threshold defaults to zero for the same reason.

Those two zeroes are the entire gap a paid cleaner is filling on a default installation, and both of them are a setting in the Storage page rather than a purchase. We think that is the most useful sentence on this page: the first-party tool leaves alone exactly the two folders most likely to be large, and it leaves them alone on purpose, because it will not delete a file you downloaded on the assumption you are finished with it.

§3The categories a cleaner adds on top are mostly caches that already cap themselves

Delivery Optimization cache, max age: 259,200 s — three days
Delivery Optimization cache, max size: 20% of available drive space
Local crash dumps: off by default · 10 files when enabled

Here is the list a third-party product adds beyond the first-party handlers, with the path each one means and what the operating system already does to it. Every row is from a Microsoft page cited in §9.

Category	Where it lives	What Windows already does
Per-user temporary files	The first of TMP, TEMP, USERPROFILE, then the Windows directory	Storage Sense deletes temporary files not in use, on by default
System temporary files	C:\Windows\SystemTemp	Separate path returned to SYSTEM processes, inaccessible to others
Recycle Bin	Per volume	Storage Sense deletes items older than 30 days
Downloads	%USERPROFILE%\Downloads	Nothing, by default. Threshold is 0 = never
Update download cache	%SystemRoot%\SoftwareDistribution\Download	Documented only as a repair step, not as maintenance
Delivery Optimization cache	%SYSTEMDRIVE% by default	Expires at three days; cache cleared automatically to stay under 20%
Local crash dumps	%LOCALAPPDATA%\CrashDumps	Feature is off by default; oldest file replaced past 10
Setup and rollback data	$Windows.~BT\Sources\Panther, %WINDIR%\Panther	Previous installation deleted ten days after upgrade
Component store	%SystemRoot%\WinSxS	StartComponentCleanup task, waiting 30 days

Two of these deserve a sentence each. Delivery Optimization is the clearest case of a self-limiting cache being sold as an accumulation: Microsoft documents a default DOMaxCacheAge of 259,200 seconds and a default DOMaxCacheSize of 20 per cent, and states that “Delivery Optimization constantly assesses the available drive space and automatically clear the cache to keep the maximum cache size under the configured percentage.” And SoftwareDistribution is documented as a repair procedure with its own caution: the manual reset renames %Systemroot%\SoftwareDistribution\Download to Download.bak rather than deleting it, and Microsoft says those steps “should only be performed at this point in the troubleshooting if you can’t resolve your Windows Update issues after following all steps”. A product that empties it weekly is running a repair on a machine that is not broken. Ledger row D-09 records that as DEPENDS, and the condition is named: whether Windows Update is actually failing.

§4Explorer reports the component store as far larger than the space it occupies, and Microsoft says so on the page

Explorer-reported size, Microsoft’s example: 4.98 GB
Actual overhead, same example: 507.18 MB
Cost of /ResetBase: 0 existing update packages remain removable

This is the mechanism that settles the argument, and it is one command. Microsoft states the cause flatly: files in WinSxS “might appear to be stored in more than one place in the operating system, but there’s usually only one copy of the file and the rest of the copies are actually hard links”, and “some tools, such as the File Explorer, determine the size of directories without taking into account that the contained files might be hard linked, which might lead you to think that the WinSxS folder takes up more disk space than it really does.” A hard link is a second directory entry pointing at one set of blocks. Adding up names does not add up bytes.

The supported way to get the real figure is in the box:

Dism.exe /Online /Cleanup-Image /AnalyzeComponentStore

It prints three quantities that most arguments about this folder collapse into one. Windows Explorer Reported Size of Component Store is the inflated one, and Microsoft annotates it: it “doesn’t factor in the use of hard links”. Actual Size of Component Store does factor them in. Shared with Windows is the part that is hard-linked into normal Windows operation and “shouldn’t be considered part of the component store overhead”. What is left — Backups and Disabled Features plus Cache and Temporary Data — is the overhead, and Microsoft works its own example: the folder “appears to be 4.98 GB, but the actual overhead … is 507.18 MB.” A tenth.

Microsoft’s own worked example. The top outline is what a directory listing reports. The lower row is the same store measured with hard links counted once: the long outline is the part shared with Windows and carried anyway, and the two solid blocks at the right are backups, disabled features, cache and temporary data. The dashed line marks shared ↔ overhead. Only the solid part is reclaimable, and it is 507.18 MB of a folder reported as 4.98 GB.

Reducing it is also in the box, and it has already happened. The StartComponentCleanup scheduled task “automatically cleans up components when the system isn’t in use”, waiting “at least 30 days after an updated component has been installed before uninstalling the previous versions”, with a one-hour timeout. Running Dism.exe /online /Cleanup-Image /StartComponentCleanup does the same work without the grace period and without the timeout. The /ResetBase switch goes further and states its own price: “All existing update packages can’t be uninstalled after this command is completed.” That is a TRADE with a number on both sides, which is what the category almost never publishes. Microsoft also warns, on the same pages, that deleting from WinSxS directly “might severely damage your system, so that your PC might not boot, and make it impossible to update.”

§5Three of these categories cost something real when you delete them

Prefetch gain, Microsoft’s own tests: 10%–80% of launch time
Layout.ini rebuild interval: 72 h of system idle
Same gain on an idle SSD: NOT ESTABLISHED — see §8

The first is any cache that exists to avoid recomputation, and the clearest example is application-launch prefetching. A Microsoft servicing blog describes the mechanism: the prefetcher watches up to the first ten seconds of a launch, writes a trace to a per-application .pf file in %systemroot%\Prefetch, and thereafter “if the .pf file exists, the kernel component of the prefetcher issues asynchronous IOs to prefetch metadata, data, and image pages described in the trace file”. Delete the traces and the next launch of each application has none, so it gets no prefetch until the trace is rebuilt. The same post publishes the size of what is at stake, which is rare enough to quote: “we saw between 10% and 80% improvement in application launch times across multiple devices, multiple applications, with multiple configurations”. It also names the condition — “the smallest gains are on SSD devices with no load at all” — and that condition is why Ledger row D-04 reads HARMFUL rather than catastrophic. Every 72 hours of idle, the same service writes launch-ordering data to %systemroot%\Prefetch\Layout.ini for a defragmenter to use. No Microsoft guidance we could find has ever recommended emptying this folder.

The second is boot and layout data, which is the same argument with a different clock, and it DEPENDS on the storage type in exactly the way the blog says: sequential issue still helps a solid-state device, and helps least when nothing else is competing for the queue.

Claim Clean system trace files by Windows Winset. Archived Windows Winset capability index, tlwinset.com/help.htm, capture 2012-08-24. Checked 2026-08-06 · Ledger row D-08

The third is anything that is evidence. Local user-mode crash dumps are the sharpest case, because Microsoft states both halves: “this feature is not enabled by default”, and when an administrator does enable it under HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows\Windows Error Reporting\LocalDumps, DumpCount defaults to 10 and “when the maximum value is exceeded, the oldest dump file in the folder will be replaced with the new dump file.” So a product reporting recovered space in %LOCALAPPDATA%\CrashDumps is reporting on a folder that is either empty or already self-limiting — NO-OP, Ledger row D-08. Where somebody has deliberately turned the feature on, the verdict changes shape rather than degree: the deletion works, and its cost is that an intermittent fault becomes undiagnosable. The same applies to $Windows.~BT\Sources\Rollback, which Microsoft describes as holding the logs, the device-install trace and, if the machine bug-checked during an upgrade, a mini-dump.

§6The category sells disk space and prices it as speed, and those are different quantities

Space, measured, in the worked example: 507.18 MB of 4.98 GB reported
Seconds recovered by deleting a cache: NOT ESTABLISHED — see §8
Microsoft criteria a scan claim can match: 3 bullets, Lack of choice

Every mechanism above is a statement about bytes. Not one of them is a statement about elapsed time, and the two do not convert into each other at any published rate. Freeing 507 MB on a 512 GB volume changes no documented behaviour of the memory manager, the file system or the scheduler. The one place the documentation runs the other way is the caches: a deleted cache costs time until it is rebuilt, and the prefetch figures in §5 are the only measured numbers on this page that touch latency at all. They are negative.

We think the honest description of the product is a disk-space feature, and that the durable version of the criticism is not that it does nothing — some of it deletes real bytes — but that it is priced against a benefit nobody in the category has published a measurement for. That is an opinion about an engineering and marketing decision, and it is labelled as one. What is not opinion is the position Microsoft takes on the alarming presentation of the figure, quoted in §1.

§7Free disk space becomes a performance problem below two documented thresholds

Free space defrag needs to finish: 15% of the volume
Page-file growth trigger: commit charge at 90% of the commit limit
Page-file growth ceiling: 3× RAM or 4 GB, and ≤ 1/8 of the volume

There are cases where disk space is genuinely the performance problem, and Microsoft publishes numbers for two of them, so the answer is a threshold rather than a sentiment. HARDWARE.

The first is defragmentation, which still matters on rotating media: “a volume must have at least 15% free space for defrag to completely and adequately defragment it… If a volume has less than 15% free space, defrag will only partially defragment it.” Microsoft names the reason — the tool uses that space as a sorting area — and names the remedy in the same paragraph, which is to delete files or move them elsewhere. That is a real, documented case where clearing space improves something measurable, and it is the one the category almost never cites, presumably because it applies to a shrinking share of machines.

The second is the commit limit, and it is the sharper of the two. “System-managed page files automatically grow up to three times the physical memory or 4 GB (whichever is larger, but no more than one-eighth of the volume size) when the system commit charge reaches 90 percent of the system commit limit. This assumes that enough free disk space is available to accommodate the growth.” Two consequences follow directly from that sentence. A volume with no free space cannot grow its page file, so the commit limit stops rising; and Microsoft states what happens next: “if the system commit charge reaches the system commit limit, the system and processes might not get committed memory. This condition can cause freezing, crashing, and other malfunctions.” The one-eighth clause also means the ceiling is a function of volume size, so a small system volume caps the page file regardless of how much RAM is installed. This is the mechanism by which a nearly full disk really does make a machine behave badly, and it is a memory-manager mechanism rather than a disk one.

Ledger row D-07 covers the tempting shortcut here, which is deleting the hibernation file. It is a TRADE: Microsoft documents that a hiberfile’s “default size cannot be smaller than 50” per cent of installed memory and that HiberFileSizePercent lives under HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Power, so the space is real — and turning hibernation off also removes Fast Startup, which on many machines is the thing that was making boots quick. Row D-10 covers the other one, reserved storage, and it is UNVERIFIED for a specific reason we repeat here: the size is the whole claim, and Microsoft publishes no size for it on either the DISM page or the support page. What Microsoft does say cuts against the framing, since reserved storage “increases the likelihood that Windows updates can be downloaded and installed without users having to free disk space”.

§8Standing: what this page establishes and what it does not

Costs in primary sources: 19 cited, all accessed 2026-08-06
Costs in unresolved rows: 2 UNVERIFIED of 10 touched
Costs we measured ourselves: 0 — we run no software

Standing — 2026-08-06 Documented The ten cleanmgr categories and the one-week temporary-file rule; the VolumeCaches/StateFlags mechanism; the Storage Sense defaults of 30 days, 0 days and 0 days; the hard-link explanation and the 4.98 GB / 507.18 MB worked example; the StartComponentCleanup 30-day wait, one-hour timeout and the /ResetBase consequence; the Delivery Optimization cache defaults; that local crash dumps are off by default and capped at 10; the ten-day deletion of the previous Windows installation; the prefetch trace mechanism and the 10–80% figure; the 15% free-space requirement for defrag; and the page-file growth rule. Inferred That the two Storage Sense zeroes are the gap a paid cleaner fills on a default installation follows from the published defaults plus the category list, not from any product we examined. That a scheduled SoftwareDistribution purge is a repair run on a working machine follows from Microsoft’s own caution about when to perform it. Our judgement That this is a disk-space feature priced against an unmeasured speed benefit; that the reclaimable-overhead arithmetic in §4 is the single most useful thing a reader can take from this subject; and that a tool unwilling to publish its definition of junk has told you something about the number. Not established Any published junk definition or measurement method from a product in this category (§1). What the archived Windows Winset module actually enumerated or deleted — the capture names the feature and no path, so the coverage question is permanently unanswerable from the record we hold (§1, §5). Any Microsoft documentation of the thumbnail or icon cache location, size cap or eviction policy; we checked the Disk Cleanup handler documentation and the Storage policy CSP and found none, so no thumbnail-cache claim appears above. A published size for reserved storage (§7, row D-10). Any independent measurement of elapsed-time change after a junk clean (§6). We also could not read ftc.gov press pages on this date — both the 2024 Restoro and Reimage release and the complaint PDF returned HTTP 403 to our fetcher — so no regulatory claim is made on this page, and the Microsoft criteria page in §1 carries that argument instead.

§9Sources & method: every claim above, with the page it came from

Costs in fetches: 19 sources, fetched 2026-08-06
Costs in vendor prose republished: 0 sentences · 2 quotes, ≤9 words each
Costs in screenshots: 0, permanently

Method, plainly: every Microsoft page below was fetched and read on 2026-08-06, and the archived vendor page was fetched from the Internet Archive on the same date at the capture URL given in both claim cards. Where a figure is a Microsoft example rather than a general fact — the 4.98 GB store, the 507.18 MB overhead — it is labelled as an example in the body. Nothing here was measured by us: this page contains no timings, no benchmarks and no first-hand account, because we do not run the software this category sells and we have never run Windows Winset.

cleanmgr (Windows Commands), ms.date 2024-01-18. Accessed 2026-08-06
Automating Disk Cleanup tool, original KB 253597 — VolumeCaches and StateFlagsNNNN. Accessed 2026-08-06
Deprecated features in the Windows client, ms.date 2025-09-09 — read in full; contains no Disk Cleanup entry. Accessed 2026-08-06
Storage Policy CSP, ms.date 2025-03-12 — Storage Sense cadence and threshold defaults. Accessed 2026-08-06
Delete your previous version of Windows — the ten-day automatic deletion. Accessed 2026-08-06
Determine the Actual Size of the WinSxS Folder — hard links, the output fields and the worked example. Accessed 2026-08-06
Clean Up the WinSxS Folder — the task, the 30-day wait, the one-hour timeout, /ResetBase. Accessed 2026-08-06
Delivery Optimization reference, ms.date 2026-05-20 — cache age, size and automatic clearing. Accessed 2026-08-06
Collecting User-Mode Dumps — LocalDumps, DumpCount, default state. Accessed 2026-08-06
Additional resources for Windows Update, ms.date 2026-02-12 — the SoftwareDistribution reset procedure and its caution. Accessed 2026-08-06
DISM Storage reserve command-line options — what reserved storage is for. Accessed 2026-08-06
Chad Duffey, What is Application Launch Prefetching?, Microsoft TechNet blog, archived on learn.microsoft.com — the trace mechanism, Layout.ini, and the 10–80% figure. Accessed 2026-08-06
defrag (Windows Commands), ms.date 2023-09-28 — the 15% free-space requirement and the SSD cadence. Accessed 2026-08-06
Introduction to the page file, ms.date 2026-02-12 — commit limit, growth rule and the free-space precondition. Accessed 2026-08-06
Powercfg command-line options — /hibernate, /size, the 50 per cent floor, HiberFileSizePercent. Accessed 2026-08-06
How Microsoft identifies malware and potentially unwanted applications, ms.date 2026-01-29 — the three Lack of choice bullets. Accessed 2026-08-06
Log files and resolving upgrade errors, ms.date 2025-04-08 — $Windows.~BT\Sources\Panther and the rollback contents. Accessed 2026-08-06
GetTempPath2W function (fileapi.h), ms.date 2025-03-11 — the environment-variable search order and C:\Windows\SystemTemp. Accessed 2026-08-06
Archived Windows Winset capability index, tlwinset.com/help.htm, Internet Archive capture of 2012-08-24 — the two quotations in the claim cards, each verified against the capture. Accessed 2026-08-06

Previous · What happens before your desktop appears — the boot path as a sequence of documented phases, so a tool claiming to shorten it has to name the phase it is shortening.

Next · Sixteen places Windows can start a program — a first-party enumeration of every autostart category, and the capability-boundary question every startup manager should have to answer in public.