Building an AI Customer Support Agent with Persistent Hindsight Memory

Customer support AI is broken. Every time a ticket closes, conventional chatbots experience total amnesia. When a loyal customer returns two weeks later with a recurring checkout failure, they are treated like a total stranger.

For the Vectorize.io Hackathon, I built ApexSupport, an intelligent customer support agent powered by the Vectorize Hindsight Triad (Retain, Recall, Reflect) and Next.js 14.

The Architecture: Retain, Recall, Reflect

1. Retain: When a customer interacts with support, resolutions and device preferences are preserved as structured nodes in a dedicated Hindsight Memory Bank.
2. Recall: When the customer returns, a multi-strategy retrieval pipeline runs in parallel:
   • Dense Semantic Similarity (Cosine distance)
   • Sparse BM25 Keyword Search
   • Causal Knowledge Graph Traversal (Entities & past solutions)
   • Temporal Recency Scoring
3. Reflect: The agent synthesizes beliefs about customer disposition and past frustrations before formulating a proactive, empathetic resolution.

Live Demo & Features

• Stateless vs. Hindsight Toggle: Direct side-by-side comparison of standard amnesic bots vs. memory-augmented AI.
• Interactive Storyboard: Walk through a 60-day customer journey from first-touch to device migration.
• Real-World Live Login: Allows any user to log in with their real name and device to test live memory retention across sessions.

Check out the live interactive demo: https://customer-support-memory-agent.vercel.app/