Author(s)
Prof. Sayali Gaikwad , Ajim Najir Shaikh, Nitin Rajkumar Kote , Sanika Jitendra Kamble
- ISSN (P): 3139-8464
- Manuscript ID: 141130
- Volume: 2
- Issue: 10
- Pages: 690–701
Subject Area: Computer Science and IT
Abstract
Large language models (LLMs) are now general-purpose components for conversation, coding, research and document work, yet they are usually delivered through separate applications that fragment workflows, duplicate context and disconnect AI assistance from the artifacts users work on. This paper reviews the technologies behind an emerging class of systems termed AI workspaces, in which conversational AI shares persistent state, authentication and data with productivity modules. Five enabling areas are examined: persistent memory, retrieval-augmented generation, agentic tool use, multimodal AI, and efficient local inference. A reference architecture is synthesized, security and privacy threats are analyzed, and an evaluation framework for workspace-level systems is proposed. The ASTRA AI project, a twelve-module unified workspace with a Next.js and Spring Boot implementation and a local Qwen3 1.7B prototype, serves as a case study that separates planned design from implemented functionality. The review identifies open problems in memory scaling, permission-aware retrieval, small-model reliability, user isolation and workflow-level benchmarks.