Vector databases store high-dimensional embeddings and retrieve them by similarity rather than exact match, making them the retrieval substrate for RAG and agent memory; the capability is now proven and ubiquitous, but by 2026 it ships natively inside every major cloud and data platform, which is a problem for the standalone category rather than for the technology.
Summary
A vector database stores items as embeddings: fixed-length numeric vectors produced by a model, where semantic similarity corresponds to geometric proximity. Instead of a query returning rows matching an exact predicate, it returns the k nearest vectors under a distance metric. Because exhaustive comparison across hundreds of millions of vectors is too slow, these systems rely on approximate nearest-neighbour index structures that trade a controlled amount of recall for large speed gains, plus filtering to combine similarity search with ordinary metadata predicates. This is the retrieval half of retrieval-augmented generation, and the persistence layer behind most agent memory designs.
The parameters that decide the outcome are therefore commercial rather than algorithmic. First, whether workloads are large or specialised enough that a purpose-built engine beats a bolt-on extension on cost, recall and latency at scale. Second, whether the operational simplicity of keeping vectors next to transactional data (one system, one consistency model, no dual-write) outweighs any such advantage for the median application. Third, whether the owners of the write path buy rather than build: in 2025 Snowflake and Databricks spent roughly $1.25B acquiring PostgreSQL-first companies (Snowflake/Crunchy Data ~$250M, Databricks/Neon ~$1B), Supabase raised a $100M Series E, and IBM acquired Confluent for about $11B at $31.00 per share cash, completing 17 March 2026 and folding the Kafka streaming write path into watsonx as a real-time data fabric for agentic AI ref.
The tension is that the standalone layer is still growing while being absorbed. The vector database market is put at about $3.73B in 2026 growing at roughly 23.5% CAGR, with an alternative series of about $2.7B in 2025 rising to about $8.9B in 2030 ref. Growth of the category and commoditisation of the capability are happening simultaneously, which is exactly the pattern that separates a durable infrastructure business from a feature.
Evidence base
Open questions
- At what scale, filter selectivity and write-churn rate does a purpose-built vector engine measurably beat pgvector/pgvectorscale or warehouse-native vector search on recall, tail latency and total cost? No supplied source contains this comparison.
- What share of the ~$3.73B 2026 vector database market is standalone vendor revenue versus vector features bundled into cloud and data platform contracts?
- Do agentic workloads change the retrieval access pattern enough (long-lived memory, high update rates, per-agent isolation) to re-open a specialist opportunity that RAG alone did not sustain?
- Does the write-path absorption pattern continue after the IBM/Confluent close, or was 2025-26 a one-off consolidation window driven by cheap targets?
Assessment drafted 2026-08-31 from up to 10 KB sources using the technology-scorecard framework; scores are a draft read pending review.