Title: Implementing Hierarchical Long-Term Semantic Memory for LinkedIn’s Recruitment AI

Abstract: As Large Language Model (LLM) agents become integral to commercial applications, the ability to deliver personalized, context-sensitive interactions has become a critical requirement. This functionality relies heavily on a robust long-term semantic memory system capable of distilling both explicit and implicit cues from complex, longitudinal behavioral data, organizing this information structurally, and facilitating rapid access. Developing industrial-scale memory systems for LLMs presents five primary obstacles: scalability, retrieval speed, privacy compliance, adaptability, and system observability. To address these issues, we present the Hierarchical Long-Term Semantic Memory (HLTM) framework. HLTM structures textual inputs into a schema-aligned memory tree that captures semantic insights across varying levels of detail. This architecture supports scalable data ingestion, privacy-conscious storage, swift retrieval, and clear data provenance. Additionally, the framework includes an adaptation mechanism designed to ensure effectiveness across a wide range of applications. Comprehensive testing on LinkedIn’s Hiring Assistant demonstrates that HLTM increases answer accuracy by over 5% and boosts retrieval F1 scores by more than 10%, while substantially optimizing the balance between query and indexing latency. The system is now fully operational within LinkedIn’s Hiring Assistant, driving essential personalization capabilities in live recruitment processes.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC