LinkedIn has built a Cognitive Memory Agent (CMA) as a shared infrastructure layer that gives AI agents persistent, stateful memory across sessions. The system organizes memory into three layers: episodic (interaction history), semantic (structured knowledge about users and entities), and procedural (learned workflows). CMA serves as a shared substrate for multi-agent coordination, reducing state duplication and enabling consistent outputs across distributed workflows. It incorporates recent context retrieval, semantic search, and memory compaction via summarization to manage storage and relevance at scale. The design addresses the fundamental statelessness of LLMs and represents a broader industry shift toward memory-driven, production-grade agentic systems.
Sort: