"The system runs smoothly for short conversations, but after around 40-50 messages, the response relevance starts dropping. The AI begins repeating phrases or giving vague answers unrelated to the context. Here's what I've already tried: Trimming the old messages to stay within token limits Summarizing older context dynamically Reducing temperature for better control Even tried vector-based context recall - still not consistent"
"I'm working in Triple Minds and I'm working on a Candy AI Clone project like this. The system runs smoothly for short conversations, but after around 40-50 messages, the response relevance starts dropping. The AI begins repeating phrases or giving vague answers unrelated to the context. It feels like the AI "forgets" the topic when the chat gets too long. Has anyone managed to maintain long conversation accuracy in their AI clone? Would love to hear what worked for you."
A Candy AI Clone project experiences high-quality responses for short conversations but degrades after roughly 40–50 messages. The model starts repeating phrases and producing vague, context-unrelated answers. Multiple mitigation strategies were attempted: trimming older messages to fit token limits, dynamically summarizing older context, lowering temperature, and using vector-based context recall; none provided consistent improvement. The behavior feels like the model "forgets" the topic as the chat lengthens. The developer seeks practical approaches and shared experiences for maintaining accuracy and coherence in long-form conversational sessions.
Read at SitePoint Forums | Web Development & Design Community
Unable to calculate read time
Collection
[
|
...
]