What you need to know
- AICore may temporarily use large amounts of memory while updating AI models in the background on your device.
- Google keeps both old and new AI models up to three days to fail during updates.
- Once the stability of the new AI model is confirmed, the memory used by AICore is automatically freed.
If you have AICore installed on your Android phone and you’ve noticed that it’s taking up an unusually large amount of memory, Google has finally explained why.
Over the past few years, companies like Google and Samsung have been implementing on-device AI technology in their phones. On most flagships Android phonesmuch of this functionality is handled by the AICore software.
AICore software is actually a system service powers the Gemini Nano on the device on Android. It enables offline and personal AI features like smart replies and notification summaries on devices like Pixel 10 Pro. But if you’ve noticed that it’s using a lot of memory, you’re not alone—some users have reported it Takes up to 11 GB.
Article continues below
Google has just updated AICore support page to explain this behavior. The company says that you “may sometimes find that this service uses more memory than expected.” According to the support page, this happens when the background AI model is updated to a new version.
To make sure everything works reliably, AICore temporarily stores both old and new versions of the AI model on your device for up to three days. This provides failover protection, allowing the system to immediately roll back if an error occurs during the update without requiring you to redownload large files.
Adoption of Android Central
I understand why Google did this, and it’s a smart fail. But 10GB+ for a background service is still wild. At the very least, there should be a way to manually limit or control this.
Google also says that this extra storage is automatically freed up once the update is confirmed to be stable, so there’s no need to take any action.
Honestly, this approach makes sense. If the app immediately deletes the old model and the update fails, you’ll have to reload several gigabytes of data and lose access to AI features in the meantime. This method makes the whole process more reliable.
