Production AI agents need two distinct context layers: retrieval (stateless lookup against a pre-built index for grounding responses in external knowledge) and memory (stateful, mutable tracking of past interactions per user). Retrieval answers 'what's in my data' while memory answers 'what happened before.' Using only one leads to agents that are either forgetful or ungrounded. Stitching two separate systems together introduces cumulative latency from extra network hops, silent freshness drift between stores, and dual-write consistency problems. These infrastructure issues manifest as context poisoning, distraction, confusion, and clash — collectively called context rot. The post argues for consolidating both retrieval and memory into a single real-time context layer, and introduces Redis Iris as a solution that combines vector search, short- and long-term memory, hybrid retrieval, and semantic caching in one engine.
