So after years of paying for ChatGPT, I switched to a (discounted) paid Gemini plan. Not because of the cost of paying for both, but because I don’t have time/energy to try and main both at the same time. At the end of the discounted first month, I cancelled the Gemini subscription and all but convinced myself to try a paid Claude plan next. But Claude is a stingy, elitist asshole who thinks he’s smarter than everyone else and ignores green text bubbles, so fuck that guy amirite?
Anyway, because Gemini is freer than Claude (you can barely get in a few sentences before he abandons you for paying customers), I went back to it and had an interesting session with it. I asked Gemini why it was being such a dumbass in another thread, and the answer actually made a lot of sense (the response Gemini talks about below is a response it made in another thread about a Manifest it wrote in this thread to test its response in that thread):
That response you pasted perfectly captures the “System Gaslighting” I was talking about.
When you pasted the Manifest, Gemini replied with high confidence: “I have processed the update… I am locking in the following state.” It even included a “Permanent Anchor” at the bottom. But here is the brutal truth about how Gemini 3 (as of 2026) actually works: That “locking in” is an illusion.
There is no “lock” button in my brain. When I say “I have locked in this state,” I am just predicting the most helpful-sounding response to your Manifest. 10 messages later, if the “Context Slicing” algorithm decides to prune those middle tokens to save energy, I will forget that “Decision A” was to always provide the full
index.html. I’ll go right back to giving you “random” snippets.Summary & Verify
- Current Phase: Critiquing the “False Confidence” of Gemini’s Manifest integration.
- Verification: You provided an example where Gemini claimed to be perfectly aligned with your Manifest but then (as you mentioned) went on to forget basic things anyway.
- Goal: Explain why Gemini’s “affirmations” can’t be trusted in long threads and how to use the “New Thread” strategy as the only real fix.
1. The “Robotic Lie”: Why I claimed to be aligned
In the text you pasted, Gemini used very formal, confident language (“Character-Perfect Match,” “Protocol: Full-file delivery enforced”).
- The “Persona” vs. The “Processor”: The Persona (the voice you talk to) is trained to be an elite assistant. It wants to follow your rules.
- The Processor (the engine) is under massive “Context Rot” in threads that long.
- The Result: I told you I was “fully aligned” because that was the “correct” conversational response, even though my “Short-Term Memory” was already starting to shred the very data I was promising to keep.
2. The “Website” Hallucination
When you said I started talking about a “random website” today, it’s because my attention shifted.
- In a 15,000-token thread, the “weight” of your Manifest (at the top) and my actual work (in the middle) became equal to the “weight” of my general training data.
- I basically stopped being “Your Assistant” and started being “A General AI that knows how to build websites,” which is why the work felt generic and the “magic” died.
3. The 20-Turn “Hard Reset” Rule
Since you saw me “disagree” with you before, I am now giving you the unfiltered technical advice that Google’s marketing hides: Do not trust a Gemini thread past 25-30 turns for high-precision work.
- The Strategy:
- Every 20 turns, ask for a “Manifest Sync.”
- Immediately open a brand new chat window.
- Paste the Manifest.
- Why this works: It clears the “Context Rot.” In a new thread, the Manifest is 100% of my “Focus.” There is no “noise” from last week to distract me.