Interview with Yang Yu: Databases Face Their Most Dramatic Role Shift in 50 Years

Since the 1970s, database evolution has centered on helping people extract value from data, first through early ETL pipelines and then by standardizing SQL as a universal interface that hides implementation details such as language choice, storage location, and table design.

Yang Yu, Vice President of R&D at KuKe Data, argues that this 50‑year data‑engineering paradigm is now breaking down. The rise of Agentic AI changes three fundamentals: the operating subject shifts from humans to agents, data types expand from purely structured to fully multimodal, and memory becomes a core database capability because agents need both short‑term session state and long‑term experiential knowledge.

The interview details the memory problem: short‑term memory requires high‑frequency reads/writes, low consistency, and a lifespan of minutes or hours, while long‑term memory demands infrequent writes, high read frequency, strong consistency, and support for diverse formats (text, audio, video, embeddings). Agents generate dynamic conversation logs, user preferences, and execution histories that must be stored and retrieved efficiently.

To support agents, the article proposes a unified "Agent‑Ready" data platform with four essential traits: (1) native multimodal storage for tables, text, images, audio, and vectors; (2) a mixed‑retrieval and execution engine that can plan and run combined vector, full‑text, and relational queries without shuffling data across separate systems; (3) a semantic layer that translates business intent into data operations for agents; and (4) robust observability and audit mechanisms—including fine‑grained permission control, data masking, access logs, and human‑in‑the‑loop approval for autonomous writes.

KuKe Data implements this vision on the PostgreSQL ecosystem: it uses Directory Table to address external object storage, supports Iceberg lake formats, adds a distributed vector engine and full‑text search, and builds a cloud‑native stateless compute layer with push‑down capabilities. The service layer exposes MCP Server and semantic APIs, while extensions such as pgvector, pgaudit, and time‑travel tables provide vector search, strong consistency, and audit trails.

Compared with competitors like Snowflake and Databricks, KuKe Data emphasizes a single, PostgreSQL‑compatible system that avoids double‑copy CDC pipelines, thereby preserving consistency and simplifying architecture. Yang Yu concludes that the shift from human‑centric to agent‑centric databases is as revolutionary as moving from abacuses to computers, demanding a complete redesign rather than a simple plugin.