Enterprise Knowledge Agents Using Vector Embeddings and Persistent Memory Architectures

Abstract

Purpose. Enterprise knowledge access fails structurally because useful organizational information is fragmented across heterogeneous systems, expressed in inconsistent vocabulary, and restricted by real security boundaries. This article examines why conventional enterprise search and short-context language model assistants fail and proposes an integrated five-component reference architecture combining semantic vector retrieval, structured persistent memory, and policy-governed access control enforced before retrieval begins.

Design/methodology/approach. The article employs a structured synthesis of peer-reviewed and preprint literature spanning information retrieval, retrieval-augmented generation (RAG), large language model (LLM)-based autonomous agents, knowledge graphs, enterprise knowledge management, and dense passage retrieval to derive architectural requirements and the reference design presented.

Findings. Five structural failure modes of conventional approaches are identified. A ten-property comparative analysis confirms no existing architectural pattern jointly addresses semantic retrieval quality, cross-session continuity, and pre-retrieval access governance simultaneously. The proposed five-component architecture — retrieval, memory, authorization, orchestration, and response layers — addresses all ten properties.

Originality/value. The article advances the position that enterprise knowledge agents must be designed as long-lived, governed organizational systems rather than stateless retrieval wrappers. The integrated architecture and access control enforcement framework fill a gap in existing literature. The ten-property architecture comparison provides the first structured evidence that the integration gap is real across all three dimensions simultaneously.

Research limitations/implications. The architecture is derived from structured literature synthesis and has not been validated against a primary empirical enterprise deployment. Future empirical case study research should measure session continuity improvement, retrieval precision under role constraints, and reduction in repeated investigative effort.