The Stack Overflow Podcast Inside LinkedIn's cognitive memory agent for agentic personalization
12 snips
Aug 25, 2026 
Praveen Bodigutla, Principal AI researcher at LinkedIn who led development of their cognitive memory agent, explains a four-layer tree-structured memory for a hiring assistant. He discusses conversation, episodic, procedural, and semantic layers. Topics include ingestion and compaction of interaction streams, storage and access-control at scale, reducing LLM calls with incremental updates, and balancing freshness, latency, and relevance.
AI Snips
Chapters
Transcript
Episode notes
Memory Agent Gives Recruiters Persistent State
- LinkedIn built a cognitive memory agent to give recruiters a persistent personalized state across hiring workflows.
- The agent manages an ingestion-retrieval lifecycle to capture expressed preferences, feedback, and role refinements and surface them later.
Four Layer Memory Captures Different Granularities
- Memory is layered: conversation, episodic, procedural, and semantic to capture recency, provenance, behavior, and aggregated preferences.
- Each layer serves different granularity: short-term chat state, temporal activities, how users act, and long-term aggregated signals.
Compress Interaction Streams At Ingestion
- Compress streaming interactions at ingestion to avoid context bloat and preserve signal.
- Identify session boundaries, subtopics and compact them in real time so retained memory stays relevant and discoverable.