Databricks Launches Lakebase Search in Beta for Agent-Native Postgres Retrieval

Databricks introduced Lakebase Search as a hybrid vector and full-text retrieval system available in beta on AWS and Azure. [1] The system uses two native Postgres extensions, lakebase_vector and lakebase_text, to let an entire agent loop rely on a single backend. [1] Agents currently operate four times more databases on Lakebase than human users do. [1]

In typical agent schemas, chunked documents and embeddings exist alongside an active conversational memory log, creating a continuous read/write loop. [1] Agents write new learnings to memory and need that data fully indexed and searchable on the next turn. [1] Vector search creates severe data bloat because documents are split into chunks that each generate a high-dimensional embedding. [1]

The introduction of Lakebase Search underscores a shift from treating search as a static read-only snapshot to a live operational database for AI agents. By integrating retrieval directly into Postgres, Databricks is targeting the memory bottlenecks associated with industry-standard vector indexes like HNSW, which the company notes are fundamentally memory-bound. This single-backend approach contrasts with deploying standalone vector databases alongside traditional relational stores, simplifying the architecture for agentic applications. What the sources don't address: specific performance benchmarks comparing Lakebase Search against established standalone vector databases or pricing details for the general availability release.