Hi everyone,
I am working on a conversational AI avatar built on Convai, configured with a Knowledge Bank containing detailed information about training courses. I am experiencing a persistent issue with context retention across conversation turns.
The problem
When a user mentions a course in one message and then asks a follow-up question without repeating the course name, the avatar fails to connect the follow-up to the previously mentioned course and responds with “I don’t have information on this topic.”
Example 1:
User: “How long is the course Sales Induction - Licence to Sell?”
Cleo: “The course is 40 hours long.”
User: “Does it have a facilitator?”
Cleo: “I don’t have information on this topic.”
Example 2:
User: “What is the certification of the course Selling GPI?”
Cleo: “The course Selling GPI is certified by CIS.”
User: “What are the contents?”
Cleo: “Could you clarify which course you are referring to?”
User: “Yes, that one.”
Cleo: “I don’t have information on this topic.”
What I have tried
I added the following instructions in the character Description field, but the issue persists:
“Always maintain the context of the conversation. If the user asks a follow-up question without specifying the subject, always refer to the topic discussed in the previous message. Never treat each message as an isolated question. If the user asks ‘and the duration?’ or ‘who is it for?’ or ‘how do I access it?’, assume they are still referring to the last course or topic mentioned. Only reset the context if the user explicitly introduces a new subject. When in doubt, always default to the last mentioned course or topic. When the user replies with a confirmation such as ‘yes’, ‘sì’, ‘ok’, interpret it as a confirmation of the last course mentioned and answer directly.”
Why I am not using Long Term Memory
I am aware of the Long Term Memory feature but I have deliberately chosen not to enable it. The avatar is used by multiple users simultaneously and I am concerned about cross-user interference — context from one user’s session influencing another user’s conversation. What I need is a session-level memory: isolated per user and reset automatically at the end of each session.
My questions
- Is there a native Convai feature (session memory, variables, conversation state) that handles this use case?
- If yes, how should it be configured?
- If no, what is the recommended approach?
Thank you in advance for any guidance.
What do you need help with? Feature request
Issue area Context Handling
Impact Blocker - I cannot continue
Character ID eafad6b8-439f-11f1-bc2f-42010a7be02c
Session ID 73d88c7c-d9d9-4603-aac6-04ebe0b4207f
Where does it happen? Convai Playground
Short summary Context not retained between turns when using Knowledge Bank
Recent character changes Added instructions in the Description field to maintain conversation context and handle follow-up questions without course name repetition.
User message or prompt “Does it have a facilitator?” (asked after the user had already mentioned the course “Sales Induction - Licence to Sell” in the previous message)
Current response “I don’t have information on this topic. Click the yellow Help box at the top right to discover the topics I can help you with.”
Expected response The avatar should retain the course name mentioned in the previous turn (“Sales Induction - Licence to Sell”) and answer: “Yes, the course Sales Induction - Licence to Sell is facilitated by a CIS Sales Facilitator.”
Steps to reproduce
- Ask the avatar a question that includes a specific course name (e.g. “How long is the course Sales Induction - Licence to Sell?”)
- Receive a correct answer from the avatar.
- Ask a follow-up question about the same course without repeating the course name (e.g. “Does it have a facilitator?”)
I checked the relevant documentation, resources, and similar forum posts before posting.
on
I confirm that I did not include API keys, tokens, passwords, secrets, private character data, private conversation logs, or customer data.
on