Skip to content

Recall on a Disk Image: What Released Builds Leave

On a disk image of a released build, Recall content is encrypted. What you can still establish: which user had Recall, when it was saving, and deletions.

Published on 9 min read

TL;DR. On a released build, a dead-box image gives you no Recall content: the snapshots and ukg.db are encrypted, the keys are protected by the TPM, bound to the user's Windows Hello identity and used inside a VBS enclave, and no public method decrypts them offline. You also need the BitLocker recovery key to read the volume at all, since Recall requires disk encryption. What the image still holds is the footprint: which profiles have a store, the files' sizes and NTFS times, the USN journal's record of writes and deletions, older copies in Volume Shadow Copies, the policy values in SOFTWARE and NTUSER.DAT, and the build number, which tells you whether the store could be readable at all. From those you can usually say who had Recall, when it was saving, and whether its history was deleted. The tool on this site reports this footprint from the files and hives, and never tries to decrypt anything.

Why the content is out of reach

Microsoft's security architecture (September 2024) and Manage Recall describe the protection:

  • Snapshots and "any associated information in the vector database are always encrypted", each snapshot with its own key.
  • The keys are protected by the TPM, tied to the user's Windows Hello Enhanced Sign-in Security identity, and used only by operations inside a VBS enclave, a memory region isolated by the hypervisor from the kernel and from administrators.
  • Data is decrypted "just in time", after the user authenticates with Windows Hello, and the authorisation times out.
  • Microsoft says other users, administrators and Microsoft cannot view a user's snapshots.

The research agrees on the storage side. Kevin Beaumont, testing the Release Preview in April 2025, found the database in the same folder, readable as a file without admin rights, but encrypted. Kaspersky wrote in October 2025 that "there are no publicly known methods to decrypt the database directly". Alexander Hagenah, whose 2026 work extracts Recall data from a live session, calls the cryptography "genuinely solid": his method waits for the user to authenticate on the running machine. None of this applies to an image.

So the honest statement in a report is that the content exists, is encrypted by design, and cannot be read from the image. This site does not describe ways around Windows Hello, the enclave or the encryption.

First, the build

The build number decides what a store can be. Read CurrentBuild and UBR under Microsoft\Windows NT\CurrentVersion in the SOFTWARE hive.

BuildWhat a store can be
24H2 (26100) pre-release builds from May and June 2024, with Recall enabled by handPossibly plain SQLite and JPEG: see the ukg.db schema
26120.2415 and later Insider builds (from November 2024)The reworked, encrypted feature
26100.3915 and later (from April 2025)The released, encrypted feature

Check the files as well as the build: a database that does not open as SQLite is not readable, whatever the build says. The hardware matters too. Without a Copilot+ PC there is no Recall store to find.

What the image still holds

EvidenceWhereWhat it can show
Store foldersUsers\*\AppData\Local\CoreAIPlatform.00\UKP\{GUID}\Which profiles had a store. Each user opts in separately
File names and sizesukg.db, -wal, -shm, ImageStore\, *.sidbWhether the store is empty or holds weeks of history
NTFS times$MFT, $STANDARD_INFORMATION and $FILE_NAMEFolder creation, the database's last write, the WAL's last write
USN journal$Extend\$UsnJrnl:$JFile creations, overwrites and deletions in the store, with times
Volume Shadow CopiesSystem Volume InformationEarlier sizes and times of the same files, still encrypted
Policy valuesSOFTWARE, each NTUSER.DATWhether Recall was allowed and whether saving snapshots was turned off. See Recall policies and settings as evidence
Other profilesUsers\*A store in an account nobody expected deserves a question

Two limits apply to all of it. Public sources do not describe how released builds write snapshots to disk, so do not read the number or names of files in ImageStore\ as a snapshot count unless your own tests on the same build support it. And SQLite writes to ukg.db when it checkpoints the WAL, not at every snapshot: the WAL's last write can be later than the database's, and neither is exactly the last snapshot time.

Opt-in has no documented registry value

The user's choice to save snapshots is made in Recall's first run or in Settings, with Windows Hello. Microsoft documents no registry value that records it, and neither does any research found for this article. The store itself is the evidence: a ukg.db that grew over weeks shows that snapshots were saved. The policy values only say what was allowed.

Reading the footprint

What you seeReasonable readingWhat else to check
No CoreAIPlatform.00 folder in any profile, no Recall policyRecall was probably never set upHardware; the USN journal for deleted store files
A store with a small ukg.db and old times onlyRecall was tried and stopped, or saving was turned offPolicy values; the dates against the user's history
A large store, last writes within minutes of the incidentRecall was saving snapshots during the incident; the content exists but is encryptedWhether the device is still available for a lawful live examination (collection)
A sharp drop in size, or mass deletions in the USN journalHistory deleted: by the user (one app, or Reset Recall), by retention, by a policy change, or removal of RecallPolicy key changes at the same time; the user's logon sessions; who was active
Writes stop at a precise timeSaving paused or turned off, the disk below 25 GB free, the user signed out, or a policy appliedFree space, logons, policy values
A store under an account created during the intrusionThe account went through Recall's opt-in with Windows HelloLogons, Windows Hello enrollment for that account

A deletion right after an intrusion is worth a close look. Hagenah reported in 2026 that code running as the user can call Recall's delete function without Windows Hello. Microsoft treats his findings as within the documented design. So a wiped store can be the user's choice, an administrator's policy, or an intruder's cleanup, and the footprint alone rarely tells them apart. Correlate with the logon sessions and with the processes running at the time.

Time windows also matter. With a maximum age of 90 days, which Microsoft's policy reference gives as the default, a store examined in November cannot hold snapshots from June, even if nobody deleted anything. A missing old snapshot is retention, not tampering.

Older copies and the WAL

Volume Shadow Copies can hold earlier versions of ukg.db and the WAL. On a released build they are encrypted too, but their sizes and times extend the footprint backwards: a store that was 2 GB a week before the incident and 40 MB after it was cut. On a readable 2024 store, a shadow copy can hold rows that were deleted since.

What the tool does with such a store

Dropped as a folder or an archive, or handed over from Disk Image Parser, an encrypted store is listed file by file with sizes and times, the database is reported as not readable as SQLite, and the policy values found in SOFTWARE and NTUSER.DAT are shown with their documented meaning. Nothing is decrypted, guessed or uploaded: the analysis runs in the browser.

Writing it up

  • State the facts: "A Recall store exists in the profile of <user> on <host> (ukg.db, <size>, last written <time> UTC). Its content is encrypted by design and cannot be read from the disk image."
  • Give the build and the source of each time ($SI, $FN, USN).
  • Separate what the footprint shows (snapshots were being saved in this period) from what it cannot (what they showed).
  • Do not write that a missing store proves Recall was never used, or that a store proves the user saw any particular thing.

FAQ

Can I decrypt a Recall store from a disk image?

No public method does. Microsoft protects the keys with the TPM, ties them to the user's Windows Hello Enhanced Sign-in Security identity and uses them only inside a VBS enclave. A disk image contains neither the TPM nor the running enclave. Kaspersky wrote in October 2025 that no public method decrypts the database directly, and the 2026 research by Alexander Hagenah works on a live session after the user unlocks Recall, not on files.

Why collect an encrypted Recall store at all?

Because its presence, size and times are evidence: which user turned Recall on, roughly when, whether it was saving snapshots during the incident, and whether the history was deleted. Keep the files too: they may matter if the device is still available for a lawful live examination.

Does a CoreAIPlatform.00 folder prove that Recall was used?

Not on its own. Public sources do not document when Windows creates the folder. A ukg.db that grew over weeks, with recent writes, is much stronger evidence that snapshots were being saved.

Does a missing store prove that Recall was never used?

No. Removing Recall, a reset or a policy that turns off saving snapshots deletes them. Check the policy keys, the USN journal and Volume Shadow Copies for traces of a store that was there.

Frequently asked questions

Can I decrypt a Recall store from a disk image?

No public method does. Microsoft protects the keys with the TPM, ties them to the user's Windows Hello Enhanced Sign-in Security identity and uses them only inside a VBS enclave. A disk image contains neither the TPM nor the running enclave. Kaspersky wrote in October 2025 that no public method decrypts the database directly, and the 2026 research by Alexander Hagenah works on a live session after the user unlocks Recall, not on files.

Why collect an encrypted Recall store at all?

Because its presence, size and times are evidence: which user turned Recall on, roughly when, whether it was saving snapshots during the incident, and whether the history was deleted. Keep the files too: they may matter if the device is still available for a lawful live examination.

Does a CoreAIPlatform.00 folder prove that Recall was used?

Not on its own. Public sources do not document when Windows creates the folder. A ukg.db that grew over weeks, with recent writes, is much stronger evidence that snapshots were being saved.

Does a missing store prove that Recall was never used?

No. Removing Recall, a reset or a policy that turns off saving snapshots deletes them. Check the policy keys, the USN journal and Volume Shadow Copies for traces of a store that was there.

Related articles

A worked example on a synthetic Recall store: alice's morning, the invoice email, the .pdf.exe opened at 09:12:07, what Recall shows and what it cannot show.
The WindowsAI policy values for Recall in SOFTWARE and NTUSER.DAT, what each proves about whether snapshots could be saved, and what no value shows.
Recall's ukg.db as public research documents it: WindowCapture events, apps, URLs, OCR text, image tokens, epoch-millisecond times, and what is still unknown.