Pinecone vs Weaviate vs Chroma: Choosing the Memory Layer for Your Agent Stack
“Pinecone vs Weaviate vs Chroma” is one of the first searches a builder runs once an agent needs to remember something past the current context window, and it is harder than it looks, because the differences that matter are not really about search quality. All three return solid nearest-neighbor results on a demo dataset. What diverges is ownership: who can cut off your access, what you can run yourself if you have to, and what the invoice looks like once an agent is calling the index thousands of times a day instead of a few dozen times in a notebook.
These three are the vector stores builders shortlist most often for agent memory: long-term recall across sessions, retrieval-augmented generation over a document set, or an agent searching its own prior work. They start from different assumptions about who owns the infrastructure. Pinecone is a managed service with no self-hosted version at all. Weaviate and Chroma are both open source, which sounds like a shared trait until you look at how each one turns "free to run" into a paid product at scale. We picked this apart while wiring memory into real agent stacks, and this is what actually differs once you are past the getting-started tutorial.
What each one actually is
Pinecone is a fully managed, serverless vector database with no open-source or self-hosted release. The only way to run it outside Pinecone's own cloud is Bring Your Own Cloud, an enterprise option that still runs Pinecone's closed binary inside your account rather than open code you can inspect or fork. In exchange, you get dense, sparse and full-text indexes, hosted embedding and reranking models through Pinecone Inference, and a built-in retrieval layer called Pinecone Assistant, all behind one API with nothing to patch, upgrade or scale yourself.
Weaviate is an open-source vector database, released under the BSD-3-Clause license, built around storing objects and vectors together so vector search, keyword search and structured filtering run as one hybrid query instead of three systems bolted together. You can self-host it free on your own infrastructure, or run it on Weaviate Cloud, the managed offering that adds replication, role-based access control, and a Query Agent that turns natural-language questions into database operations.
Chroma is the vector store that shipped inside a huge share of early LangChain and LlamaIndex tutorials, thanks to an embedded, single-file design that made it the easiest way to add retrieval to a prototype. It is Apache 2.0 licensed and still free to self-host, either as a single Python process for local development or a distributed Rust-based deployment for production. Chroma Cloud is the newer, fully serverless hosted version, and the company now positions itself as search infrastructure for AI rather than only a vector database, with Sync and Agent products alongside the core database.
The comparison, builder's cut
| Pinecone | Weaviate | Chroma | |
|---|---|---|---|
| Center of gravity | Fully managed, zero-ops | Open source, hybrid search native | Open source, usage-first serverless |
| License / self-host | Proprietary, cloud only (BYOC for enterprise) | BSD-3-Clause, free to self-host | Apache 2.0, free to self-host |
| Where it shines | Teams that want zero infrastructure ownership | Hybrid search plus the option to self-host later | Prototyping locally, same API through to production |
| Weak spot | No self-host path if you ever need one | More moving parts if you skip the managed cloud | Youngest of the three at serverless production scale |
| Pricing floor (checked 2026-07-17) | Standard from $50/mo minimum | Flex from $45/mo minimum | Starter $0/mo, usage-based, no minimum |
| Pick it when | The vector store should disappear into the platform | You need hybrid search or may self-host later | You want one API from prototype to production |
Pricing, verified this week
Vector database pricing moves fast enough that any number older than a few months is a guess, so here is what each vendor's own pricing page showed when we checked on 2026-07-17.
Pinecone's pricing page lists four tiers: a free Starter plan capped at 2GB storage with no monthly minimum, a $20-a-month flat Builder tier for solo developers, a Standard tier with a $50-a-month minimum billed pay-as-you-go beyond that floor, and Enterprise at a $500-a-month minimum. Pinecone's cost documentation confirms the same four figures and notes the Standard and Enterprise minimums are a floor you grow into, not a charge stacked on top of usage.
Weaviate Cloud's pricing page lists three tiers: a permanently free plan capped at 100,000 objects and 10GB disk, a pay-as-you-go Flex tier from $45 a month, and a prepaid Premium tier from $400 a month covering shared or dedicated deployment. These are current figures from a 2026 pricing restructure that renamed the old Serverless and Enterprise Cloud tiers, so treat any comparison post quoting different numbers as stale rather than wrong.
Chroma Cloud's pricing page and its usage documentation show no monthly minimum at all on the Starter tier: $0 a month plus usage, with $5 in free credits to start, then $2.50 per GiB written, $0.33 per GiB stored monthly, and $0.0075 per TiB queried plus $0.09 per GiB returned. The Team tier is $250 a month with $100 of that included as usage credit, built for production workloads that want SOC 2 and direct support. Chroma is the only one of the three with a genuine no-floor path from free prototype into paid production; Pinecone and Weaviate both gate real production traffic behind a monthly minimum, $50 and $45 respectively, once usage clears their free tiers.
Open source vs proprietary: what that actually buys you
The license question is not academic once an agent is in production. If Pinecone changes its pricing or your workload needs a deployment shape it does not offer, your options are renegotiate or migrate every embedding elsewhere; there is no source code to fork and run yourself. Weaviate's BSD-3-Clause license and Chroma's Apache 2.0 license both mean the core database is yours to run forever at zero license cost, even without touching either company's managed cloud. That freedom is not free in the other sense: self-hosting swaps a monthly bill for an on-call rotation, since index tuning, backups and capacity planning become your team's job instead of the vendor's. Most teams under real load end up paying one of these three eventually anyway, just later, with a clearer idea of what the managed version is worth to them.
Lock-in and migration risk
None of the three looks like a near-term graveyard entry. Pinecone and Weaviate are both venture-backed with paying enterprise customers, and Chroma's move from a Python side project into a distributed Rust deployment plus a funded cloud product is not an investment a team abandons quietly. The lock-in that matters here is quieter than shutdown risk: Pinecone Assistant and Inference are convenient precisely because they are integrated, so the retrieval logic living inside them does not travel if you leave. Weaviate's Query Agent and Chroma's Agent product create the same gravity in miniature. Our advice does not change by vendor: keep chunking and embedding logic outside any store's proprietary features where you reasonably can, keep raw source text somewhere you control, and treat the vector index as a rebuildable cache, not a system of record. Re-embedding a corpus costs time and money. Losing the source documents it was built from is the failure that actually hurts.
How to choose in five questions
- Do you want to think about infrastructure at all? If not, Pinecone's zero-ops model is worth its pricing floor.
- Do you need hybrid, keyword-plus-vector search out of the box, or might you self-host later without a rewrite? Weaviate's native hybrid search and open license cover both cases at once.
- Are you still prototyping and genuinely unsure of scale? Chroma's no-minimum Starter tier and local-first roots let you build against the same API from laptop through to production traffic.
- Does your retrieval pipeline already lean on a vendor's built-in embedding, reranking or agent tooling? Price that migration cost into the decision now, not after you have shipped.
- Still tied? Load a real slice of your own corpus, at your actual document size and query pattern, into two candidates for a week. The pricing calculators on all three vendor sites are illustrative, not a substitute for your own numbers.
Frequently asked questions
Is Pinecone open source?
No. Pinecone is a fully managed, proprietary vector database with no self-hosted release; even its Bring Your Own Cloud option for enterprise customers runs Pinecone's closed binary inside your account rather than open code you can inspect. Weaviate and Chroma are both open source and free to self-host.
Can I self-host Weaviate or Chroma for free?
Yes. Weaviate is BSD-3-Clause licensed and Chroma is Apache 2.0 licensed, and both can run entirely on your own infrastructure at no license cost. You take on the operational work of running them yourself, which is exactly what their managed cloud products charge to remove.
Which of the three is cheapest to start with?
Chroma's Starter tier has no monthly minimum at all, just usage plus $5 in free credits, per trychroma.com/pricing checked 2026-07-17. Pinecone's free Starter plan also has no minimum but caps storage at 2GB. Weaviate's free tier is capped at 100,000 objects, and its first paid tier starts at $45 a month once you outgrow it.
Do I need a dedicated vector database for agent memory?
Not always. At small scale, an in-memory index or a Postgres extension like pgvector can cover agent memory if you already run Postgres. Once an agent needs to recall across long histories, large document sets, or multiple sessions with low query latency, teams typically converge on one of these three, or a comparable dedicated vector store.
This layer of the stack changes its pricing pages faster than it changes its technology, which is exactly why we checked each vendor's page directly today instead of trusting a comparison post from last quarter. We publish the running version of this comparison, plus two composition notes and one graveyard entry, in the Monthly Stack Report. If you are wiring agent memory this year, that is the cheapest insurance we know of.