Hi everyone, I built an AI chatbot in Bubble using Anthropic Cloud. The chatbot answers correctly, but it’s not conversational.
For example:
- User asks a question >>> AI answers perfectly
- User asks a follow-up question >>> AI acts like it forgot the previous message/context & gives a generic answer like “Hi, how can I help you?”, etc.
- User asks a question. >>> AI answers and, at the bottom, asks the user a follow-up question. When the user answers the follow-up, AI forgets the entire conversation and gives a generic answer like “Hi, how can I help you? etc”
It feels like only the latest user message is being sent to the model, not the conversation history.
Database/workflows seem connected correctly, and messages are being saved.
What should I check first? Is this usually a backend workflow issue, API payload issue, or conversation history formatting problem?
Appreciate it if you could give me some hints so that I know what to work on / which sections to work on. 
You’ll need to include screenshots of your API Connector setup, and associated workflows, and then someone will be able to look and identify the issue 
It may not be clear, but “memory” is just a list of messages that you send back to the LLM.
SO when the user is ‘chatting’ with AI, each API call to the LLM needs to contain both sides of messages. The Agent Responses and the User Messages all in a list sent back via the API call.
Once you get that working, you can look into compacting etc, but this should get you going!
Yes, you’re right. I needed to share screenshots. I could get results faster. 
With a hint I got here, I figured it out. Thank you!
Thank you! This was helpful. That’s true. Both sides needed to be sent, which was not happening in my setup. I corrected it & worked. Thanks!
If you’re okay with adding an additional layer then I’d suggest sending API request to an n8n WF webhook. Its AI Agent node has a simple memory tool where you can set chat’s unique id/slug as unique identifier.
If you don’t want to send a list of old messages in the chat via API, then this is an alternative option
Thank you very much! This was a very interesting workaround. The issue was resolved, actually, but I’ll test this out, too. I loved it.
