matt-wisdom/bcthack

RecAgent is a consumer agent framework built for a hackathon that simulates how a real person might research and respond to products. It models cognitive memory in three layers: a fast sensory layer for processing raw product and environment data, a short term layer for managing the current session context, and a long term layer backed by a vector database that stores past behaviors and insights, with a mathematical forgetting mechanism so older memories fade over time. The system takes a persona description, for example a price conscious student from Lagos who loves photography, and uses that profile alongside real time economic data such as inflation rates and exchange rates specific to the persona's country to reason about whether a product fits that person's actual purchasing situation. It can search a local product catalog offline or use a web search tool to find products in real time. The framework exposes two main API endpoints. One generates a product review from the perspective of the described persona. The other generates personalized product recommendations based on the persona's interests, drawing from product datasets including Amazon, Jumia, and a local store catalog. You would use this if you are building or experimenting with AI agents that simulate consumer behavior, for example to generate realistic synthetic reviews, test how different personas respond to product recommendations, or study how economic context affects purchasing decisions. The backend is a containerized FastAPI application using LangGraph for agent reasoning, ChromaDB for vector database memory, and a language model for complex reasoning. It is written in Python and runs via Docker Compose.