Unofficial archive. This is a read-only, community-run archive of the Apple Support Community. It is not affiliated with or endorsed by Apple Inc.

Idea for Apple Intelligence memory consolidation on iOS

I have an idea for improving AI memory in Apple Intelligence on iOS 27 and later while preserving user privacy.


The concept is “memory consolidation.” Instead of allowing approved memories to accumulate separately, Apple Intelligence could periodically analyze existing memory during idle periods, such as while an iPhone is charging, and identify redundancy, relationships, and information that can be represented more efficiently.


For example, memories containing a user’s name, preferred name, birthday, location, and preferences could be consolidated into a smaller representation without losing meaning.


A safe system could include three states: Locked, which prevents changes; Optimizable, which allows consolidation or reorganization; and Pending Deletion, which keeps old versions temporarily recoverable.


This could run asynchronously on-device during idle or charging periods. It would optimize only information already authorized for retention, reducing storage and retrieval overhead while maintaining user control and privacy.

7 replies

First read the Unsolicited Idea Submission policy by Apple, which is similar to those of other companies. Yes, there are people who attempt to sue companies for implementation of a feature that they claim was their idea, so the policies are pretty common and necessary.

Legal - Unsolicited Idea Submission Policy - Apple


If you are still compelled to send Feedback on any feature in iOS, you would do that here:

Feedback - iPhone - Apple


The typical approach would be through Acquisition of a company that has developed a technology that Apple is interested in. That requires much more than an idea, it would actually be a proof of concept with the technology currently in use where Apple can see and verify the benefits.


That’s more commonly called data deduplication, and it’s a very common technique.


There are tradeoffs arising with deduplication, including performance and privacy.


Techniques of data deduplication were taught some forty years ago, and probably further back. This was with systems featuring 32 KW or 16 KB or less memory, meaning de-duplication, hash tables, so-called “bumming code”, and other techniques for reducing usage were all common.


I expect a forest of patents exist here, too.


Async processing add can issues other interlocking with.

Thanks for the feedback. I wasn’t familiar with data deduplication or the earlier techniques you mentioned, so that’s actually useful information for me. The Apple policy point is useful too. What I’m really interested in is how those existing concepts could potentially be applied to Apple Intelligence’s memory system—essentially giving Siri a way to organize and consolidate the memories it already has permission to retain so it can reference them more efficiently. The underlying mechanism could be whatever Apple determines works best; the idea is the improved memory organization itself.

wjones30309 wrote:
Thanks for the feedback. I wasn’t familiar with data deduplication or the earlier techniques you mentioned, so that’s actually useful information for me. The Apple policy point is useful too. What I’m really interested in is how those existing concepts could potentially be applied to Apple Intelligence’s memory system—essentially giving Siri a way to organize and consolidate the memories it already has permission to retain so it can reference them more efficiently. The underlying mechanism could be whatever Apple determines works best; the idea is the improved memory organization itself.

There’s a Harvard online course on computing and related topics, including data structures. Maybe that helps?


Harvard CS50x:

https://www.edx.org/cs50


Here’s an intro to data structures:

https://cs50.harvard.edu/x/weeks/5/


There’s a Harvard CS50 AI course available, as well:

https://www.edx.org/learn/artificial-intelligence/harvard-university-cs50-s-introduction-to-artificial-intelligence-with-python


As for topics and implementations of information storage and retrieval, that might involve a key-value store, a hash table of some sort, a linked list, or any of an assortment of other data structures, with some implementation choices dependent on other factors including scale and scope and frequency.


And the whether and the how of data de-dup is implemented can vary, as can oft-related overhead from garbage collection.


Hardware details matter, too. De-dup isn’t as helpful if the only copy of the data is not local, for instance. Do a remote access or cache miss or storage access or memory compression too often, and the expected savings can evaporate. (I don’t think any current Apple silicon Mac is NUMA, so there's that.)


Trade-offs are everywhere.

abdelrahim96 wrote:
I can not found Apple Intelligence in m’y phone iPhone 17 pro Max modèle CH/A

You don't have it on that model. No iPhones intended for China distribution are allowed to have that capability.


In the future, instead of hijacking a post from another user about something that is completely unrelated to the question they asked, you should create your own question using the link on the top of the page that says Ask the Community. That way you give the original poster the respect they deserve to have their issue addressed.

I can not found Apple Intelligence in m’y phone iPhone 17 pro Max modèle CH/A

😂😂😂No offense taken on my part; I understand how they could have made that mistake because I had difficulty figuring out the posting/reply mechanics myself.