AI agents lack persistent memory between calls, causing them to forget prior context. Semantic memory search solves this by storing facts as vector embeddings outside the model and retrieving them by meaning rather than exact keywords. Key challenges include retrieval speed (latency multiplies across multi-step agents) and data freshness (stale memories can silently degrade responses — one study found 60.5% of top recalled entries contained outdated information). Hybrid search combining vector similarity with full-text ranking covers each method's blind spots. Redis Iris is presented as an integrated platform offering vector search, semantic caching (Redis LangCache), and durable agent memory (Redis Agent Memory) with a two-tier session/long-term design. Production results cited include LinkedIn's hiring assistant achieving 5–10 percentage-point reduction in negative feedback, and semantic caching delivering up to 73% lower inference costs in high-repetition workloads.
