LinkedIn's hiring agent runs four distinct memory layers inside a 10-20% share of the response latency budget

Stack Overflow Blog·medium signalXBlueskyLinkedInCopy link

In a Stack Overflow podcast published 2026-08-25, LinkedIn Principal AI Researcher Praveen Bodigutla describes the memory system behind the hiring assistant: conversational memory for the current session, episodic memory with temporal querying and provenance back to source interactions, procedural memory capturing how each recruiter makes trade-offs, and semantic memory aggregating preferences across sessions and product surfaces. The team moved off GraphRAG to a tree-structured memory for faster incremental updates, and holds the whole memory agent to 10-20% of total response latency using parallel planning, selective LLM calls, and prefix caching. A three-tier evaluation checks entity preservation and citations to prevent context pollution.