Today I noticed there’s now three ‘model’ options available to choose from, Fast, Thinking and Pro. Over the past two weeks I’ve been having major issues related to context retention and the model being too aggressive in pruning info even within 10-20 prompts. When I saw the new model options, I thought maybe that’s solved, but if anything, it’s just verified and codified the shortcomings of 3.0 more.
Asking Gemini directly about the differences, it verified that Thinking allows many prompts per day (100-500), but it uses ‘Dynamic’ thinking, which means it may prune/summarize earlier chat turns to maintain speed in as little as 10-20 turns. On the other hand, Pro does not do that, and maintains it’s coherence across long context windows, but…has a 10-20 prompts/day limit.
Honestly, neither of these options solves the problems many of us have that use Gemini as a tool for long conversations / back and forth for brainstorming business or creative ideas, troubleshooting coding or technical problems or really anything else that requires a ‘partner’ that you can work with on something from start to finish.
I understand there is a computational cost related to being able to have a large context window and be able to have many prompts per day, but the current setup basically offers the worst of both worlds. Below is the exact quote from Gemini when I asked it to summarize these challenges:
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”).
I’d pay more a month (within reason) to have access to the best of both worlds for context retention and prompts/day. You all had it working in 2.5, so not sure why 3.0 is such a drastic step backwards related to this issue. Heck, you could even just give us access to 2.5 again…
On Pro Subscription my Gemini loses context after a a while.
I’ts kinda weird to have this happen when on subscription.
I understand when Free tiers have limitations and do nto always deliver, but i though if i pay for somethign it delivers what is promised in the desciption, 1mil token context window.. i loaded 9k words and Gemini forgot it aprox 20 queries later. This must be a bug, because that doesn’t align with the desciption in the Subscription Plan.
Back in 2.0 to 2.5 ( March to May 2025) i was able to write and analyze a Novel with Gemini that has 90k Words. The AI even HÆllucinates when it shouldn’t, and apologizes not even knowing why it happened. The only way it to work with Gemini then is to re-feed the files it needs, but if you do that more often it sprinkles the output with [cite:numers213] …which has turned a running joke between me and the AI…
I Hope we see a fix for this soon, OpenAI is always profiting from these slips, i know Google can do better.
I have encountered this problem too, but before seeing this thread I thought it was just me.
This is a huge problem and I’m getting ready to start a new project, so I need to know (and I bet thousands of others do too) RIGHT NOW whether this is going to be fixed, or at least offer us the ability to go back to 2.5 Pro.
If I can’t get a definitive answer on this within 12 hours (yes, normally that would be unreasonable, but Google has known about this problem for weeks now), I’m going to have to vote with my dollars and move to another AI.
Sadly, I ended up moving over to another model (ChatYouKnowWho) as it doesn’t seem to have this issue. Context window isn’t as long, but what’s the use of having a context window of 1M tokens when the model loses track of what you were talking about 4-5K tokens in?
I really don’t want to move as I really like 2.5 Pro and absolutely loved 3.0 for the first week or so, but it’s not a viable solution for anything other than simple use cases now for me. I appreciate the fact how powerful Gemini is, and the massive strides they’ve made, but it currently feels like you’re working with a Rocket scientist who has a really bad case of ADHD.
The steps are very simple: Give it too much information. While it claims a 1M token context, it can’t keep track of 50,000.
Have a long conversation. Or upload a spec doc. I spent a lot of time writing a 40 page spec. I uploaded it as a PDF. Very quickly it became obvious that Gemini 3 Pro had no idea what was in my document. Sure, it got the broadest strokes, but all important nuance was lost. The database schema was a joke. Most functions were missing. It was so bad that in the end I threw it away and started over with Claude.
A 40 page document, plus some conversation is maybe 30,000 - 40,000 tokens. Even if it was 100,000, why tout a 1 million context window when it can’t pay attention to 1/10th of the information?
As far as I can tell, the 1M token context is marketing hype. Try to get it to do much with 1/20th of that and you get junk.
I have the same issue.. I had moved many conversations from ChatGPT to Gemini, but my conversations are getting pruned, so it doesn’t have any history or context. One of my chats deleted all conversations older than 1 week or so, for example.
Same issues here. Had signed up for Gemini Pro with the Google Workspace Sub with the hopes it could be my assistant during long projects. I had crucial information within a chat with my custom gem and about 30 queries later it couldnt recall what we previously talking about.