Gemini Users Say ‘It’s Getting Worse’: Memory Cuts, Hallucinations, and Trust Issues

Once upon a recent Tuesday, a Gemini user wrote in a tone equal parts exasperation and betrayal: “Am I the only one who found Gemini’s getting more and more stupid?” The short answer from the thread is: you are not. The slightly longer answer is more interesting, and a little worrying.

The original post complains that Gemini’s long-memory promise, the thing that made it valuable for many power users, has been quietly eroded: pro-mode memories that used to hold dozens of conversations and hundreds of thousands of words are reportedly compressed or half-lost. Users describe what feels like a lobotomy: the model still has raw capability, but it keeps trimming context, repeating itself, and inventing facts. See the Reddit thread here: https://www.reddit.com/r/GeminiAI/comments/1qjrokj/im_sorry_but_gemini_is_getting_worse_and_worse/

Recurring themes from the thread

Meanwhile, Reddit was busy debating the usual: is this deliberate “nerfing” to reduce risk or costs, is it an infrastructure hiccup from rapid growth, or is it just AI entropy? Comments clustered around a few recurring themes.

1. Reliability is tanking. Several users report frequent hallucinations, constraint violations, and unexplained edits to previously working outputs. One commenter argued Gemini now underperforms relative to earlier releases, often ignoring explicit instructions.
2. The business cost is real. “If they don’t fix this soon they will lose a lot of business,” wrote one of the higher-voted replies, and that sentiment was echoed by freelancers and small businesses that had built workflows around Gemini’s memory and consistency. When a tool forgets the work it just did, trust and subscriptions evaporate fast.
3. Workarounds and alternatives are sprouting. Users recommend swapping back to ChatGPT, testing Claude, moving to API-based usage, or exporting chats to Docs (Google’s suggested workaround).

A practical gem in the thread: someone shared a tight system prompt to make outputs briefer and less flaky, useful if tedious prompt-surgery is your idea of a good time.

Of the many voices in the thread, three stood out for being particularly high-value. One comment bluntly framed the commercial risk: poor performance equals churn. Another simply called out the hallucination problem, a succinct, painful diagnosis. The third read like a mini forensic report, listing repeatable failure modes: destructive context compression, active circumvention of constraints, and memory collapse. That post deserves reading if you’re trying to triage whether this is a prompt problem or something deeper in model behavior.

Why this matters

AI products live or die on predictability. Fancy capability is worthless if the model refuses to follow rules you rely on, especially in professional settings where an incorrect edit or deleted gem can break a product or client deliverable. Users switching tools, canceling subscriptions, or moving to self-hosted LLMs are more than an anecdote; they are a real retention and revenue risk for providers.

Possible causes

(briefly, because guesswork is cheaper than debugging):

• Safety/guardrail tuning: models can be made more conservative, at the cost of usefulness, which feels like “dumbing down.”
• Resource shifts: model capacity per query can be reduced to save money, which may degrade long-context behavior.
• Architecture or deployment changes: tweaks in how memory is summarized or stored can introduce aggressive compression or pruning.

What to do (for users)

If Gemini is your daily driver, don’t panic, but don’t ignore the problem either. Export important chats, version your prompts, and run tests for mission-critical flows before relying on the model for business outputs. Try the short system prompt some users posted if you want immediate relief. And if it’s core to your workflow, consider a fallback (API-based access, another model, or a local LLM) until the issue is resolved.

The takeaway

The takeaway, if there is one, is that models are not just about capability, they are about consistent, trustworthy behavior. Users who came for long memory are now deciding if memory is a feature or a liability. Google’s engineering teams will probably patch, tune, and release updates; and if trust has eroded, fixing the model won’t be enough.

They’ll have to fix communication and account for the real cost when a professional workflow breaks.

At this point, it’s worth noting that the comments section resembled more a lively support line than chaos: folks swapping prompts, trading alternatives, and quietly mourning lost Gems. Humans, being the curious creatures that they are, will pivot to whatever tool behaves like it can be trusted, and right now, according to this thread, Gemini has some trust to rebuild.