> the user does face provider switching costs (time and effort), which serves as a mini-moat for status quo provider.
When a provider gets memory working well, I expect them to use this to be a huge moat - ie. they won't let you migrate the memories, because rather than being human readable words they'll be unintelligible vectors.
I imagine they'll do the same via API so that the network has a memory of all previous requests for the same user.
Is memory all that useful for using these LLMs? I’ve found that I mostly use them for discrete tasks - helping me learn a specific thing, build a specific project, debug a specific piece of code, and once it’s done I’d actually prefer it to forget that thing instead of keeping it around forever.
Hell, “just open a new chat and start over” is an important tool in the toolbox when using these models. I can’t imagine a more frustrating experience than opening a new chat to try something from scratch only for it to reply based on the previous prompt that I messed up.
Unintelligible vectors might be hard to transfer from one of their older models to one of their newer models - so I think the human readable words will remain a bit of a narrow waist in this space for the immediate future at least.
Gemini 2.5 Pro already uses encrypted 'thoughts' in the context window which are not visible to the user. They might be English words, or might be some other internal state vector.
When a provider gets memory working well, I expect them to use this to be a huge moat - ie. they won't let you migrate the memories, because rather than being human readable words they'll be unintelligible vectors.
I imagine they'll do the same via API so that the network has a memory of all previous requests for the same user.