Redis and AI Agent Memory with Andrew Brookins

What It Covers

Andrew Brookins from Redis explains how AI agents require sophisticated memory systems spanning short-term conversations, long-term facts, and semantic retrieval to maintain continuity across stateless LLM interactions in production environments.

Key Questions Answered

• •Memory Architecture: AI agents need three distinct memory layers: working memory for current message history, summarization systems to compact past conversations when context limits are reached, and extracted long-term facts stored separately for retrieval across sessions.
• •Hybrid Search Strategy: Vector search alone proves insufficient for production agents. Effective systems combine semantic vector search for exploratory tasks with keyword exact-match search for specific queries, fusing results based on task requirements rather than predetermined patterns.
• •Temporal Fact Management: Store timestamps with all extracted memories to enable recency-based retrieval and decay. Episodic memories bound to specific times require different handling than stable semantic facts, with time-ordered queries preventing outdated information from overriding current preferences.
• •World Model Gap: Current agents fail at predicting environment state changes despite extensive context engineering. Production agents interacting with infrastructure or dynamic systems need reinforcement learning approaches beyond language prediction to generalize across different scenarios and environments successfully.