Context Retention Issue: Merlin Forgets Instructions Within the Same Conversation
S
Souhail Ads
I've noticed that Merlin doesn't seem to retain information within the same conversation or project. Unlike ChatGPT, which learns from what I provide and continues to deliver outputs based on that context, with Merlin, I often find myself repeating the same instructions to get the desired results. It's a bit frustrating and disrupts the workflow.
This was tested using Claude 3.5 Sonnet. I hope this can be addressed soon. Thank you!
S
Siddhartha
Dana Doron Yume Souhail Ads can you please share the chat IDs
Y
Yume
Siddhartha No I am sorry I can’t. Next time it happens I can because I didn’t set out to save chat IDs about this issue, I just was making a comment agreeing with the other person… It mostly happens in instructional instances where it either gets lost or doesn’t remember context as it answers the next question in relation to the prior context. It gets confused with follow up questions in the process. Or starts providing mixed information that doesn’t retain the prior context.
S
Siddhartha
Yume I see, thanks regardless! We want to see some of these chats and see where the LLM context is messing up.
Y
Yume
Agree. I am constantly refeeding it context.
S
Siddhartha
Dana Doron can you please share the chatID where you faced this issue?
D
Dana Doron
Siddhartha I assume it's the number string in the URL? This is what I have: d913aa53-40d9-4d57-b8f7-e9b56696fbc6
D
Dana Doron
You'll see that toward the end, I was copying and pasting the full instructions with each request (though I was repeating the same process with new content each time). It was very annoying but the only way to get accurate results.
D
Dana Doron
Siddhartha Just making sure you saw my reply
S
Siddhartha
Hey Dana Doron, we have switched some thing up this past week for pro users, it should be much better in capturing context.
D
Dana Doron
I am having the same problem with Claude 3.7 sonnet. Incredibly frustrating. If it does something not quite the way I want it to after I've provided full instructions, I can't just tell it what to fix as it will have forgotten all my previous instructions. Nor can I tell it to do the same operation again (e.g., revise a new document with the same instructions). I have to repeat
all
instructions with every new request or correction.