# Vd · Vector Store — element 14 of 58

> Memory your agents can search. Turns documents into answerable knowledge.

- **Group:** 4 · Knowledge & Memory
- **Necessity:** Optional
- **Price band:** $ · under $30/mo
- **Maturity:** Stable
- **Edition:** v2026.Q3 · verified 2026-09-13

## Leading tools (v2026.Q3)

- **pgvector** — the default start
- **Turbopuffer** — serverless scale, object-storage
- **Pinecone** — managed classic
- **Qdrant** — open-source, filter-heavy
- **Chroma** — agent-native developer favorite

## Our take

Start with pgvector inside the database you already run. Graduate only when scale forces you to.

## Combines with

Mm, Re, Kw


## The top 5 — deep dossier (verified 2026-09-13)

pgvector, unless scale has already said otherwise — it lives inside the Postgres you're probably running, costs nothing, and the 2024–26 production record is littered with teams that cut $2,000+/mo dedicated-vector-DB bills to ~$200 by moving home to it. Turbopuffer is the graduation path: object-storage-native, ~10x cheaper at scale, and the store behind Cursor, Notion, and Anthropic — $100M annualized revenue on under $1M raised. Pinecone remains the managed classic when procurement wants a pure-play vendor with every compliance checkbox. Qdrant wins for self-hosted, filter-heavy hybrid search at 100M+ vectors; Chroma when developer experience and agent-native retrieval matter more than raw scale. Hanging over all of them: AWS S3 Vectors went GA in December 2025 at 2B vectors per index, and every incumbent database now ships vector search — the standalone category is being squeezed into the billion-scale niche.

1. **pgvector** (open source (Andrew Kane + community)) — Free, PostgreSQL-licensed · runs in any Postgres — RDS, Supabase (free tier), Neon, Cloud SQL. Best for: Every startup that already runs Postgres and holds under ~10–50M vectors — which is roughly 95% of them. Why: One extension turns the database you already operate into your vector store — same backups, same transactions, joins against your real data, zero new vendor. v0.8.2 (Feb 25, 2026) keeps shipping: iterative index scans, half-precision and sparse vectors, binary quantization. Documented migrations from Pinecone to pgvector at 1/10th the monthly cost are the category's defining production story. Watch: Vanilla HNSW strains past ~50M vectors — you'll want pgvectorscale (DiskANN) or a graduation plan. Index build memory and write-heavy churn need real Postgres operational skill; 'free' is not 'zero-ops'. [https://github.com/pgvector/pgvector](https://github.com/pgvector/pgvector)
2. **Turbopuffer** (turbopuffer, Inc.) — Usage-based · Launch $16/mo min · Scale $256/mo min (HIPAA BAA, SSO) · Enterprise ≥$4,096/mo (single-tenant, BYOC). Best for: Multi-tenant AI products with millions of per-user namespaces — the exact shape of agent memory and per-customer document search. Why: Search built natively on object storage with an NVMe/memory cache: ~10x cheaper economics, sub-10ms p50 hot queries, 1T+ documents and 250M+ namespaces in production. The customer list is the 2026 AI stack itself — Cursor (its largest, 1T+ code chunks), Notion (10B+ vectors, migrated off Pinecone Oct 2024), Anthropic, Linear, Harvey. Sacra puts it at $100M annualized revenue by March 2026 on under $1M of primary capital raised — the most efficient company in this table, possibly in any table. Watch: Cold namespaces pay object-storage latency (hundreds of ms) on first hit. Closed source, no self-host below Enterprise BYOC, and a small team carrying very large customers. Revenue figure is single-source (Sacra). [https://turbopuffer.com](https://turbopuffer.com)
3. **Pinecone** (Pinecone Systems) — Starter free (2GB) · Builder $20/mo flat · Standard $50/mo min + usage · Enterprise $500/mo min. Best for: Teams that want a zero-ops, pure-play managed vector database with mature enterprise controls and don't want to think about infrastructure at all. Why: The category's brand name, still: 5,000+ customers, serverless architecture, integrated inference/embedding endpoints, SOC 2/HIPAA, BYOC. The 2025 reset — founder Edo Liberty stepping to Chief Scientist, three-time founder Ash Ashutosh taking CEO in September 2025, and a new $20/mo Builder tier — is a deliberate push down-market and toward 'knowledgeable AI' platform positioning. Watch: The cautionary tale of the category: flagship logos (Notion) left over cost, mid-2025 press cycles covered sale-exploration chatter (the new CEO says acquisition is 'definitely not on the table'), and it has raised nothing since 2023's $100M at $750M. At startup scale you are paying a premium for convenience pgvector gives you free. [https://www.pinecone.io](https://www.pinecone.io)
4. **Qdrant** (Qdrant (Berlin)) — OSS free (Apache-2.0, Rust) · Cloud free 1GB cluster · Standard usage-based · Premium/Private Cloud custom. Best for: Self-hosters and regulated teams who need heavy metadata filtering, hybrid search, and full control from laptop to 1B+ vectors. Why: The strongest open-source pure-play in 2026: Rust core, best-in-class filtered search, quantization that cuts memory 4–32x, and hybrid/multi-vector support. A $50M Series B (Mar 12, 2026, Advance Venture Partners; $87.8M total) explicitly aimed at 'agentic retrieval' keeps it independently funded while rivals consolidate. TripAdvisor runs 1B+ vectors on it. Watch: Cloud pricing beyond the free tier is quote-driven — less transparent than Turbopuffer or Chroma. Feature surface (embeddings, rerankers, agents) is expanding fast; watch for focus dilution. [https://qdrant.tech](https://qdrant.tech)
5. **Chroma** (Chroma (San Francisco)) — OSS free (Apache-2.0) · Cloud usage-based: $0.33/GiB-mo storage, $2.50/GiB writes · Starter $0 + $5 credit · Team $250/mo. Best for: AI engineers who want retrieval running in five minutes — local in a notebook today, serverless Chroma Cloud with the same API tomorrow. Why: The developer-experience pick and the default vector store scaffolded by AI coding agents: 28k GitHub stars, single-node simplicity, vector + full-text + hybrid in one API. Chroma Cloud (2025, Rust-rewritten, object-storage-backed) added a genuinely cheap transparent usage model, and the team's retrieval research (context-rot, generative benchmarking) earns real credibility. Watch: Thinnest capitalization of the five ($18M seed, Apr 2023 — no announced round since) and the least enterprise track record at billion scale. Single-node OSS deployments hit walls that force the Cloud migration decision. [https://www.trychroma.com](https://www.trychroma.com)

### How to choose
- If You already run Postgres and hold under ~10M vectors (you almost certainly do) → pgvector. One extension, zero new vendors, joins against your real data. Revisit only when p95 latency or index-build pain says so.
- If You're building a multi-tenant AI product — per-user memory, per-customer document search — or your dedicated-DB bill crossed four figures → Turbopuffer. Namespaces-on-object-storage is purpose-built for that shape, and it's the migration Notion and Cursor already validated.
- If Procurement wants a managed pure-play with every compliance box ticked, and nobody wants to run infrastructure → Pinecone Standard ($50/mo min) — accepting that you're paying the convenience premium the rest of this page is about avoiding.
- If You need self-hosting, data residency, or heavy metadata-filtered hybrid search at 100M+ vectors → Qdrant OSS, with their cloud or hybrid cloud when ops outgrows you. Milvus/Zilliz if you're genuinely at billions.
- If Your vectors are cold, batch-queried, and already in AWS — archives, logs, agent episodic memory → S3 Vectors: no minimums, up to ~90% cheaper, sub-second cold queries. Pair it with a hot tier; it is not a low-latency serving store.

### The field (19 more)

Weaviate, Milvus / Zilliz Cloud, LanceDB, pgvectorscale, Amazon S3 Vectors, MongoDB Atlas Vector Search, Elasticsearch / OpenSearch, Redis (Vector Sets / Query Engine), Vespa, Supabase (pgvector hosting), Neon (pgvector serverless) (acquired), Astra DB / DataStax (acquired), Vectara, Marqo, Faiss, Deep Lake (fading), MyScale (fading), Annoy (dead), pg_embedding (dead)

Full dossier data: https://elems.ai/e/vd.json

---
Source: [elems.ai](https://elems.ai/e/vd.html) — the periodic table of the AI-led startup. Data: https://elems.ai/elements.json (CC BY 4.0, cite elems.ai).