LinkedIn has developed a cognitive memory agent designed to provide deep personalization for agentic workflows, specifically for their hiring assistant tool. Moving away from traditional GraphRAG, the team implemented a multi-layered memory architecture to manage the full lifecycle of user preferences and context. This architecture consists of four distinct layers: conversational memory for immediate interactions, episodic memory for temporal querying and provenance, procedural memory to capture user task execution patterns, and semantic memory to aggregate long-term preferences across different product surfaces. To maintain efficiency and prevent context bloat, the system utilizes an ingestion service and an offline consolidation job that performs deduplication, conflict resolution, and information augmentation. The memory is organized in a tree-structured hierarchy—linking individual project preferences to broader recruiter and cohort levels—which allows for rapid incremental updates and improved information discoverability. This approach enables the agent to handle dynamic, non-standard queries by performing temporal, episodic-based retrieval. By treating memory as a structured, queryable platform rather than a simple vector database, LinkedIn aims to increase the signal-to-noise ratio in agentic interactions, reducing user friction and improving the helpfulness of personalized recommendations.