A few weeks ago, Microsoft held an event where it introduced Copilot+ PCs. They stand out from the competition. Snapdragon X Elite and X Plus chipsets that sport a powerful NPU capable of delivering 40 TOPS (trillion operations per second). To showcase the NPU's capabilities, the software giant also introduced Recall – an app that automatically takes snapshots and creates a timeline of your PC's daily workflow. This timeline can then be searched for text that appears directly in the snapshots, or perhaps more impressively, one can search using simple visual descriptions. Thanks to the NPU, all indexing and data processing can be done locally.
The recall quickly became the talk of the town for various reasons. Some see it as a game changer, allowing people to return to a previous activity with unprecedented ease. Others see it as one. Privacy risk. After all, Recall creates a handy database of your computer usage history and habits, supercharged with screenshots.
This leads to a big question: Is there a way to test Recall on existing hardware? While the Copilot+ PCs undoubtedly look very capable, the price tag can be a tough pill to swallow — and as of this writing, they're not yet available.
Currently, only one specific build of Windows 11 supports Recall: 26100.712, which is available in the Release Preview channel and Downloadable from the UUP dump If you need a clean install. While versions of 24H2 exist in both the dev and canary channels, they do not include the necessary OS-level files.
There were two pieces of the puzzle I needed to solve to make Recall work:
- Preventing hardware requirements Check to make the PC recall-compliant.
- Installing “AI components” (machine learning models) used to process text and visual content found in snapshots
After tracking down the hardware requirements and figuring out which Windows component performs it, I came up with a plan: build a custom component that shows that all the requirements are in order and Send all checks to him. Once this requirement was met, a new immediate action appeared in the list.
This tells us that by default Recall is disabled and the user has to go out of their way to enable it in the Settings app.
Changing the toggle from off to on lights up the “Open Recall” button in the flyout and lets me access the main interface. This is where the aforementioned AI components come into play.
This flow expects to dynamically download components hosted on Microsoft's servers, however at the time of writing the files have not yet been published, resulting in an error.
Fortunately, as part of the release of Windows 11 version 24H2 to OEMs, Microsoft published the relevant components on partner channels, which many developers have access to, although currently only in the Arm64 flavor. Word quickly spread among Recall fans and the models were soon uploaded for all to see.
Oddly enough, the official OEM package comes with no installer, just some metadata and a few dozen files that sport cryptic filenames. Thankfully they weren't too difficult to realize and after a bit of file labeling I had the models registered and installed.
Another reboot, a few minutes of patience to allow the ML models to initialize for the first time, and I was done. Memory is fully activated.
Recall performance is quite good even on the modest specs of my Samsung Galaxy Book2 Go 5G laptop.
- Snapdragon 7c+ Gen 3 (4x Cortex-A78 @ 2.40 GHz, 4x Cortex-A55 @ 1.50 GHz
- 3.4 GB usable RAM
- 256 GB SSD
Snapshot creation was quick, the “screen ray” process where snapshots are processed for text and visual content took less than a second, and cleaning up the timeline was easy. The only time you notice a performance drop is when you search through snapshots. Even though the snapshots have already been analyzed and their data is stored in the database, it still takes ~6 seconds to show you the results.
So, does all this mean that the NPU requirement is a hoax? No, not at all. Just because a feature is able to function without an NPU doesn't mean it can't use it to better distribute the workload and provide a higher performance and more satisfying experience. A dedicated GPU is a good analogy. Sure, you can get by with just one integrated, but once you're dedicated the difference is definitely noticeable.
An interesting observation is that current generation Arm64 SoCs also come with NPUs – just because you can't see it in Task Manager doesn't mean it's not there. To decide how best to process the data, the Recall experience package has a component called NPUDetect that can tell what generation of NPU your PC has. While running various tests I was able to collect data for these SoCs:
- Snapdragon 7c+ Gen 3
- Snapdragon Compute Platform (8cx Gen 3, Windows Dev Kit 2023)
- Ampere Ultra
All of these are detected as NPU Generation 2 SoCs and AI Workload Manager is able to use them. I also inadvertently tested the NPU-less path when working with virtual machines. Out of the box, Recall is not possible in the Arm64 VM. The detection process recognizes the host SoC as the one with the NPU, but due to VM isolation, attempting to use it results in a crash of Recall's background processes. By patching the detection process to treat the VM as if there was no NPU I was able to get Recall to work there as well. The performance seems comparable to how the bare metal Snapdragon 7c+ Gen 3 performs, which is quite decent.
If you own Arm64 hardware and want to give Recall a spin, the tooling I've created along the way is now public. You can download it from This GitHub repository. Be sure to follow the guide you see on the landing page so you don't miss out on any bets.
If you don't have Arm64 hardware but still want to try Recall as soon as possible, there is a Emulation Guide available. Keep in mind that modern hardware emulation comes at a cost of time and a lot of computing resources, as well as a bit of a learning curve since it involves some command-line heavy lifting. However, the option is there if you want to get your hands on some Virtual Arm 64 hardware first.
Once you have all the files ready, you can simply run the installer from a command prompt, watch the automatic qualification process, and voilà. A reboot later, Recall should be ready to use.
Even if you don't plan on running Recall permanently on your main computer, this method provides a good way to learn more about this innovative but controversial feature.