Gemini is too rigid (Glassblower effect)

[glass grove]

It’s like working with hot glass: it can take any shape during the first few prompts, but once it cools, it’s stuck. If the conversation shifts, Gemini keeps forcing those early concepts back in, trying to make that initial "shape" relevant to new topics where it doesn't belong.

It feels heavily optimized for one-shot prompts where every chat is disposable. For every new topic, you basically have to build context from scratch.

This wouldn't be a problem if the memory system worked as intended. But right now, it feels completely out of my control. It claims to save things, yet when I open a new instance, it’s still relying on the same old stuff I saved months ago—nothing new actually sticks.

The dream of a "Universal Assistant" feels very far away. The memory system just feels like two extra layers of "System Instructions" with cool names—a gimmicky solution rather than real innovation. It’s essentially just a static table of "If/Then" logic. Even worse, when it finally decides to use that table, it circles back to the first problem: trying to force-fit everything to my profession or favorite series, even when I’m talking about something totally unrelated.

TL;DR
My suggestion on that big rant, is:

either; the problem is how the Gemini App itself surrounding the models feed it the context, and it needs to be changed;

or give even more clear instructions on what the "Personal context" means. I'm actively trying to use them, but they seem to either degrade the outputs or they are completely ignored in certain scenarios.

(Or the problem is that the models still unable to discern nuance)

Example: I have a instruction for Gemini to avoid finishing their responses with the generic "follow up questions" that plagues AI. And then it keeps doing them.

But then randomly at some point in the conversation it remembers the instructions and stop doing them. But that's totally up to the model to decide when to use those tools. instead of being the default, as I would assume was the intention?

And then on how hard it is to use the Personal context or understand how it was supposed to work, i don't know if it is just the model gaslighting me or what, but, i had the conversation right now about this same critique here. and the analogy of "glassblower" i used.

asked the model to save it.

It said that I had to do that manually in the "Instructions".

But i want it to be available as part of the "Past Chats with Gemini" which is a feature I have enabled.

It then says it saved it! and it will be available to new instances.

I open a new instance, and ask about this analogy. It remembers all of the old stuff i talked about with it like months ago. But nothing new, i'm stuck with this fixated memory about like 6 or 7 topics that i shared at some point, and it can't get anything new.

And when it decides to remember those things, it keeps bringing them up on totally unrelated topics.

It's a total mess!

I'm open to talk more about this issue, if anybody here wants more details, i definitely got more details!

[scarlet solstice]

Hey @glass grove I like this analogy! I will be sure to pass this suggestion along.