Context Engineering for LLM Apps: What Worked, What Didn't

Ricardo Ferreira's LLM assistant kept giving bad answers, so he switched to a newer model with a bigger context window. It didn't help, and it cost more. What fixed it was Context Engineering: memory layers, reranking, summarization and semantic caching. Ricardo Ferreira has spent 25+ years building distributed systems and databases. To stress-test an open-source agent memory server built on Redis, he put an LLM behind an Alexa skill, which must answer within 8 seconds, and then used it with his family every day. This talk covers every context failure he hit along the way and how he fixed each one. Recorded at QCon AI Boston 2026. QCon conferences are practitioner-driven events where senior engineers share real production experience. In this talk, you will learn: - Why upgrading to a larger model doesn't fix context poisoning, distraction, confusion or rot - How to combine short-term session memory with long-term vector memory, including multi-tenant post-filtering - How query compression, few-shot examples and reranking make answers more selective - Why context that grows without limit drives up token cost, and how token windows and summarization keep it flat - How semantic caching removed more than half of redundant LLM calls, and why you must recalibrate thresholds whenever you change models 🔗 Transcript & slides available on InfoQ: https://bit.ly/3Tz3lLn #ContextEngineering #LLM #AIEngineering #RAG #SemanticCaching