This is seriously embarrassing for a model this advanced. I’ve been working with Gemini 3 Pro Preview in AI Studio, and I’m running into a massive issue that makes long conversations impossible.
The biggest problem is this insane repetition bug. It happens when the chat gets just a bit long: the model will suddenly ignore my current prompt and repeat an answer it gave way earlier in the conversation, word-for-word. We’re talking about a topic we finished discussing ages ago, and it just hallucinates that we’re back there and copy-pastes its old response. It completely ignores the message JUST SENT, which is totally insane to me.
This also happens in the Gemini App (with pro subscription). I really would have reported the issue via the app but I am sure no one reads any of those.
Seems like we are in the age of trust me bro benchmarks.
I believe I’ve experienced what you’re referring to on a few occasions.
I find myself doing more initial brainstorming and development of the idea and plan of something before I get to AI studio in regular Gemini app models with various settings.
I’ve had particular success creating a custom gem specifying the persona, there is an incredibly large context window for the task specificity of the gem customization. At the moment you’re able to use various features such as deep research and canvas with this custom gem. I found this to be incredibly valuable as a starting point. Specifically using the deep research to find additional info for the project that I’m unclear on, and then doing an initial build so to speak with canvas.
This may potentially save on the chat length issue you feel may be related to the problem.
This behavior sounds like a context window handling issue rather than a prompt misunderstanding. In long chats, Gemini 3.0 Pro may lose track of the latest turn if the conversation exceeds the model’s effective token limit, causing it to re-surface earlier completions. In AI Studio, try trimming older messages or summarizing prior context before sending new prompts so the active turn stays within the token budget. If this is reproducible, capture the full conversation transcript and share it via the “Send feedback” option in AI Studio or through the Gemini API issue tracker so engineering can review. For production use, you can implement server-side conversation management that prunes or condenses history before sending to the API.
I have had this happen also, but not the entire answer. It answers the beginning of my question with correct information, and then by the end, it is quoting its old answer (old as in 1 turn old, over the course of less than a few minutes) word for word. It was unclear to me if it was doing this on purpose to try and be thorough, or if it was a mistake, but it certainly isn’t offering any value by sending the same tokens I just read. If it is doing it on purpose, it should make that clear, by saying “Keep in mind, that as I just told you…” or even a link that says “and this still applies” and jumps you up to the previous answer so that it takes a few tokens, not a few hundred or thousand. Many ways to show it’s not a bug, but since none of them were done, I assume it is.
I agree, it also likes boiling down Nuances and simplifying complex character interactions. The model is smarter but it also lacks Emotional Intelligence and has lazy context handling compared to 2.5 pro. Its gonna need a few more tweaks before it’s ready to be in-production.
I have this problem constantly beginning with 2.5 and it continues in 3.0. I’m not even having long conversations and it’s doing this. I can be 10 short messages into a conversation and then it decides to stop being useful but just repeatedly giving me what it’s previously given me and ignoring everything I’m saying, I swear it’s trolling me! It’s especially egregious when trying to work with nano banana.
Yes, this is a contextual loop failure from window context length, and you may start to see a recursive loop from media files as well, where media files will get stuck and start building up. You won’t be able to see it on your end, but if you go to Google Takeout and download the responses, it’s a good way to check because sometimes it will seem like the window should not be full, but it’s getting filled up from the media files. I had 30+ files I couldn’t see from my end, but when I was sending a prompt, Gemini was registering a massive amount of data with every prompt.
But the Google GEM solution works better now, actually, because you can pull more than 10 PDFs from your email and make a PDF summary of key points every day of your history, back up your chat sessions, and reload them every day. I have Gemini make a summary and tell it to send to Google Keep, and I manually save the whole chat but only inject the summary unless it’s something specific I’m looking for, because there is no need for the AI to keep your entire history to pull every day. Also, this will make for faster response times. Basically a manual RAG system; if you want to automate it, you can use Tasker or just build it in AI studio; both are fun to experiment with.
Another handy thing to do is tell it after every 5 responses to make a time stamp of date/time so when you need to look back, it’s easier, especially with Google Takeout; it’s a lot easier to correlate times of when something goes wrong, and oddly enough, every 5 responses, it’s like an automatic reassessment so it doesn’t lose track, but if your window is full, it’s archived your old data and can start overwriting data.
Good luck; let me know how it works out from your end.
It’s not a failure, it’s ‘by design’. I’ve been having the same issue and dove into it a bit with Gemini today and basically got told this.
The Conflict: Stability vs. Volume*
Feature
Thinking (Standard Pro)
Pro (Deep Think Mode)
Context Behavior
Uses Dynamic Thinking; may prune or summarize earlier chat turns to maintain speed.
Uses Deep Thinking; cross-references the entire window more thoroughly to prevent drift.
Stability
Reports of “context drift” after 10–20 turns; can ignore established rules or recall details incorrectly.
Highly stable; designed specifically for coherence in long-form, multi-step planning.
Daily Limits
100–500 prompts/day (varies by plan); serves as your primary high-volume tool.
~10–20 prompts/day; strictly limited due to extreme computational intensity.
Core Challenges for Long-Term Projects
The “Thinking” Memory Gap
User feedback indicates that the standard Thinking mode in Gemini 3.0 can feel less reliable for long chats than the previous 2.5 Pro version. It often struggles with “complex logical retrieval” as conversations lengthen, sometimes claiming ignorance of information visible just a few scrolls up.
The “Pro” Usage Wall
While Pro (Deep Think) solves these amnesia issues by utilizing a deeper reasoning tree, its extremely low daily cap (as few as 10 prompts for some tiers) makes it impossible to use as a primary interface for active, turn-based troubleshooting.
Instruction Adherence Decay
In long sessions, models may experience “drift,” where they revert to core training (e.g., being helpful by rewriting text) rather than following your specific project constraints (e.g., “do not rewrite”).
My guess is 3.0 was way more successful than they initially thought. There were getting hammered compute wise, so they decided to neuter everything so you can either have lots of daily interactions, but very short context memory, or lots of context memory, but very few daily interaction. The fact that this really appears to be by design implies that we’re not going to see a ‘fix’ anytime soon.
Ah ok cool. The original post was tagged for ai studio just checking. Did you check the Google takeout that I described in my previous post? I know it can cause it if media files get caught in the system. I’ve had it 3 times now and it’s been an issue since 2.5 pro came out. Also your issue is all with in a single conversation window?
No, unfortunately, it’s happening across nearly all of my ‘longer’ threads. Whether they have been live since day one of 3.0 or I start new, basically they start losing context pretty quickly. I really think this just is a design choice Google has had to make to limit computation. Lot of memory/few daily prompts or lot of prompts/frequent purging of memory.
That’s only for the App though I’ve been experiencing nuance problems in Gemini 3.0 pro even on AI Studio. The model’s personality handling and nuance handling seems one dimensional and it will often ignore instructions. It’s not related to the recent update, it’s been a problem since Gemini 3.0 pro experimental released. I’m hoping the release version is actually better than 2.5 pro before I switch to it.
What do you mean “switch to it”? It might be labeled as “experimental” in the name, but in a Pro subscription, it’s what you get. I see no way to access 2.5.
I hate to lose the good things about v3, but I’d prefer to go back to 2.5 at this point, if possible.
Just in case I’m misunderstanding the reply. Below here is the model selection menu be sure you are in AI studio not Gemini Mobile app. This is the original posts tag.
Sweet it’s a must! Imagine building a web-based app from a description and seeing a working mock-up in minutes. Anything you can imagine, and learn to code at the same time. You can ask it to make it in steps so you can learn, or build the whole thing for you instantly.
The Conflict: Stability vs. Volume*
Core Challenges for Long-Term Projects
James3,
Have you checked out AI Studio yet? If not, do so and message me if you need assistance with anything. Lots of fun stuff to build and learn.