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

Short summary

Developer built ChronoScholar, a knowledge-graph agent detecting contradictions in scientific literature using Cognee. Strict TDD with mocked dependencies masked five critical integration failures: config API removal, graph engine changes, multi-tenant scoping, file locking, and template parsing. Lesson: AI memory systems require integration tests against real dependencies in CI, not just mocked unit tests.

• •ChronoScholar detects contradictions in scientific papers via Cognee knowledge graph: 494 entities, 1,025 edges, F1=1.0 contradiction detection across 10 arXiv papers
• •Five integration failures emerged in production that mocked unit tests missed: Cognee config API removal, graph engine API changes, ENABLE_BACKEND_ACCESS_CONTROL multi-tenant scoping, LadybugDB file locking on concurrent access, Python str.strip() greedy character removal
• •Mocked unit tests fail for AI systems because they cannot detect upstream library behavior changes; recommend adding integration tests on minimal real corpus with real API keys in CI

Generated with AI, which can make mistakes.