I Built a Research Memory Agent with Cognee—Five API Breaks, One Knowledge Graph, Seven Days

I applied strict TDD to build a Cognee knowledge graph agent — here is every upstream API break I hit and how I fixed it

What I Built

ChronoScholar is a temporally-aware research memory agent that ingests arXiv papers into a Cognee knowledge graph and detects when stored scientific beliefs are contradicted by incoming literature. Built for the WeMakeDevs x Cognee hackathon.

The core problem: standard RAG returns answers from all stored papers with equal confidence. An agent that ingested Paper A in January and Paper B in March — where B refutes A's central claim — surfaces both indefinitely. Fixing this is not a retrieval problem. It requires persistent, typed memory with explicit contradiction awareness.

The system ingests papers via the arXiv API, builds a typed knowledge graph using cognee.add() and cognee.cognify(), classifies paper pairs with Gemini 2.5 Flash, and synthesizes cross-paper answers using Cognee's GRAPH_COMPLETION search mode.