What this example shows
This example focuses on memory orchestration, not just retrieval. Specifically, it demonstrates how to:- Share the same memory cluster across multiple agents
- Combine MemoryModel based customer support with autonomous insight extraction
- Persist structured commercial signals during conversations
- Separate who talks from who observes
Architecture overview
The system is composed of three main components, all connected to the same MemoryModel cluster:1. Customer Support Agent
- Answers user questions using Retrieval-Augmented Generation (RAG)
- Retrieves information from product documentation and user profiles
- Stores long-term user preferences and contextual data
2. Spy Node (SalesInsight)
- Runs silently in parallel with the conversation
- Never answers the user
- Observes messages and extracts structured sales signals, such as:
- implied needs
- sentiment
- potential upsell opportunities
- Writes structured data directly into memory
3. Operator Dashboard
- Reads only the structured memories generated by the Spy Node
- Displays real-time sales insights to a human operator
- Scores leads based on estimated conversion probability
Memory nodes used
The example uses multiple memory nodes, defined explicitly by the developer:- AppKnowledge
Stores official documentation and product manuals ingested via batch upload. - UserProfile
Stores persistent user-level information such as preferences, constraints, and past interactions. - SalesInsight
A structured node designed to capture commercial signals inferred from conversations.
nodes.txt and created explicitly when configuring the memory cluster.
This highlights a core concept of MemoryModel:you decide what is memory-worthy, not the system.
Running the example
The full source code is available as an open-source project: 👉 FinCare AI — GitHub repositoryhttps://github.com/MemoryModelRepo/CustomareCare-Upselling-finance-app-public.git To run the example locally:
- Create a MemoryModel cluster
- Define the nodes using
nodes.txt - Ingest documentation into the memory layer
- Run the customer and operator applications in parallel