I Built a Customer Support Agent That Remembers With Hindsight

Customer support agents often struggle with repeating information to customers who have already provided context about their issues. Priya, a customer experiencing Wi-Fi drops, had previously reported the same problem with her Netgear Nighthawk R7000 router. Restarting the router and updating firmware had temporary fixes, but switching Wi-Fi channels provided a more lasting solution. The challenge is to create a support agent that can remember past issues without reiterating failed troubleshooting steps.

RecallAI is a customer support agent designed to use Hindsight for long-term memory and Groq for generating responses. The key is not just storing old conversations, but retrieving the right facts at the right time, keeping customers isolated from one another, and using previous outcomes to avoid repeating failed troubleshooting steps. A stateless support agent, like most LLM conversations, doesn't have access to a customer's long-term history, so it defaults to generic troubleshooting steps.

The RecallAI architecture separates conversation context from customer memory. Short-term conversation context keeps the most recent four chat turns, while long-term customer memory stores information that remains relevant across conversations and even after the application restarts. This distinction prevents the system from sending an ever-growing transcript to the model, as only the current problem, a small recent conversation window, and relevant memories are sent to the LLM.

Customer isolation is a critical aspect of RecallAI. A deterministic Hindsight bank is created for each customer, ensuring that Priya's router information doesn't accidentally appear in Rahul's conversation. The memory service performs both retention and recall within this isolated memory bank, providing a clear boundary for each customer's memory.

Hindsight stores facts, not just transcripts. For every support exchange, it records the customer's message and the agent's response. Hindsight processes this retained information and extracts useful facts, determining what matters for the application. Retrieval of relevant memories happens before the LLM generates a response, ensuring that the agent can personalize answers without repeating failed fixes or stating history that appears in customer memory.