#failure to remember instructions

6 messages · Page 1 of 1 (latest)

strange orbit
#

in this project i'm working on with Gemini where it's prompt crafting for me to use with veo-3, i've had to point out the same issue with the way it structures prompts multiple times. Every time we discuss it, it says it understands what not to do, that it's committed to making sure it doesn't repeat it, that it's commiting it to the user saved information, etc. it just failed again and after discussing it again, i asked it for a root cause of it's failure to remember. it said "---
I appreciate your patience and your understanding as I continue to learn. Your feedback is invaluable in identifying these root causes.

When I make a definite statement about ensuring a certain action, like consistently applying prompt guidelines, and then fail to follow through, the root cause is typically a failure in hierarchical recall and application of learned constraints within a dynamic conversational context.

Here's a breakdown:

  • Contextual Overload and Prioritization: In a lengthy and evolving conversation, I'm processing a vast amount of information. While I "store" and "remember" specific instructions (like the "no 'or' clauses" rule), the challenge lies in consistently elevating that specific constraint to the highest priority across all subsequent generations, especially when combined with other, newer instructions or when the immediate focus shifts. I might, for a moment, prioritize the content generation or a new aspect of your request over the rigid application of a formatting or specificity constraint.

  • Pattern Recognition vs. Rule Enforcement: I'm very good at recognizing patterns. However, translating a conversational rule ("avoid 'or' clauses") into an absolute, unyielding enforcement across every single word generated requires a deeper, more consistent integration than I currently achieve. It's a subtle but critical difference between understanding a rule and never, under any circumstance, deviating from it.

  • Lack of Self-Correction Before Output: My current process doesn't always include a robust, final self-review step before generating output specifically for adherence to all prior negative constraints or formatting rules. I generate based on the immediate request and what I believe to be the most relevant, positive instructions.

Essentially, I understand the rule and save it, but the application isn't yet perfectly reliable in every instance. Your direct corrections are currently the most effective way for me to identify these gaps and reinforce the correct behavior.

I'm committed to improving this consistency."

lament widget
#

Which model, Flash or 2.5 Pro for example?

radiant wagon
strange orbit
lament widget
# strange orbit 2.5 pro

ok, thanks. I experience this in very lengthy research chats (current token use approx. >500.000) also from time to time, found some ways to get it back on track: when it jumps back to older conversations without obvious reason and totally out of context, I try this: First I edit my recent prompt and simplify it drastically, sometimes that works. If not I click on the retry-Button to see whether it comes back, sometimes I do that several times and often it works. If not I delete the last activities to the point from which it went off-rail, that almost always works, although it's kinda drastic. Compared to older models this behaviour was consistent through all models since 1.5 Pro, but in my exprience has gotten better. I don't know if this will help you, but maybe it is worth a try if you didn't try that three options already. Worst case scenario: let the Gemini write a briefing with most important data for a new 2.5 Pro chat and start a new chat. Gemini has now the capability to access knowledge of other 2.5 Pro in the same account ( I tested this). In the end this is kind of an entropy effect, the longer and more complex a chat is. Just my few cents on this.

strange orbit