#2656-AI Processing & Memory Clarification Request

Could you please provide brief explanations (with an example or screenshot where possible) for the following? This will help us create better AI training for our clients.

1) Incoming Message Flow

Please explain the end-to-end processing flow for a user message (e.g., message → memory → knowledge retrieval → AI instructions → response).

2) Default AI Instructions

Which instructions are always sent with every AI request (tone, language, restrictions, escalation rules, etc.)? Are these permanent or retrieved dynamically?

3) Conversation Memory

How does memory work? What is stored, for how long, and how does the AI avoid asking for information the user has already provided?

4) Vector-Based Knowledge Retrieval

How are uploaded files chunked, how are relevant sections selected, how are conflicts handled, and can we see which file/chunk was used to generate a response?

5) Knowledge Retrieval Priority

What is the retrieval hierarchy when multiple knowledge sources exist (System Prompt, FAQs, Uploaded Files, Google Sheets, Websites, APIs, etc.)?

6) Multi-File Knowledge Base

How does the system handle multiple uploaded files? Can files have categories or priorities? If a file is replaced, is its previous knowledge completely removed?

Reason:

We're currently seeing cases where the AI repeats questions, ignores previously shared information, or doesn't consistently follow instructions. Understanding the retrieval and memory flow will help us structure our training content more effectively.

Dear User,

Thank you for your questions.

1. Incoming Message Flow: Message → Conversation history → Knowledge retrieval (if enabled) → AI instructions → AI response.
2. Default AI Instructions: The configured system prompt (tone, language, rules, etc.) is included with every AI request.
3. Conversation Memory: Recent conversation history is used to maintain context and reduce repeated questions.
4. Vector-Based Knowledge Retrieval: Uploaded files are split into chunks, and the most relevant chunks are retrieved based on the user's query.
5. Knowledge Retrieval Priority: Typically: System Prompt → Conversation History → Retrieved Knowledge → AI Model Knowledge.
6. Multi-File Knowledge Base: Multiple files are indexed separately, and the most relevant content is retrieved. Replacing a file requires re-indexing to remove old knowledge.

For more accurate responses, we recommend using a higher-capability AI model. If you are using your own API key, you can select a more advanced model to achieve better response quality and instruction following.

Thank you.

Yes there are cases where the AI repeats questions, ignores previously shared information.

Also the contextual memory limits to 10 now, previously I was using 30 for one of my client why is it reduced and I see options until 10 only

We are facing the same issue after the recent update. Previously, we were able to configure the contextual memory (up to 30), but now it is limited to 10 only. This is causing AI agents to lose context, repeat questions, and enter conversation loops. We have already received similar feedback from 19-20 clients. Kindly consider restoring the configurable contextual memory option, as it was very useful for handling longer conversation flows.