Chroma Joins Replit: Chroma Cloud Winds Down in 60 Days, ChromaDB Stays Apache 2.0

Chroma is joining Replit, and anyone running production workloads on Chroma Cloud has about 60 days to move. In a post dated October 23, 2026, Chroma CEO Jeff Huber wrote that the company "has signed a definitive agreement to join Replit" to build "the context layer for the self-driving company." The same post says Chroma is "responsibly winding down" its existing cloud service over the next 60 days, while open-source Chroma continues.

This matters beyond one vendor. ChromaDB is one of the most widely used open-source vector databases for retrieval-augmented generation and agent memory, and a managed service going away on a short clock is the kind of change that breaks someone's weekend. Below: what the primary source says, what it leaves open, and a migration checklist.

TL;DR: what builders are asking

What Chroma actually announced

The announcement on Chroma's blog tells the company's story in four beats. Chroma began in 2022 working on active learning for curating training data. After ChatGPT arrived, it pivoted to inference and "the problem of context for agents." It released ChromaDB the following month, and Huber says at one point it estimated it had over 95 percent of the downloads for vector search. And in mid-2026 it started Foundation, a product for self-improving context for organizations.

Customers named in the post include SpaceXAI, Mintlify, Krea, Paramount, Retool and CoreWeave. The post also cites its technical reports on context rot and Context-1 and deep dives on its Fission transactions, object-storage-native retrieval and the wal3 write-ahead log on S3.

On the why, Huber says Replit's CEO Amjad Masad and President Michele Catasta were early investors in Chroma and "it became increasingly obvious" the two could do more together. Catasta is quoted: "We first met Jeff and the team in early 2023 and were immediately struck by the talent and deep thinking about what an AI infrastructure layer would look like."

What happens to Chroma Cloud

The sentence that matters operationally: "In this transition over the next 60 days, we are responsibly winding down our existing cloud service." The post links a Chroma Cloud wind-down timeline for customers affected. We did not find exact dates in the announcement itself, so open that page for the date that applies to your account.

AlphaSignal's coverage describes the move as a shutdown with a 60-day window, and adds general migration advice: export data, select a replacement deployment, preserve collection IDs, embedding dimensions and models, metadata, filters and access controls, then test retrieval quality and latency before switching reads and writes. That advice is sensible, though it comes from a secondary source rather than from Chroma.

What stays: open-source ChromaDB

Chroma says open-source Chroma "remains strong at 6M downloads a month and just shy of 30k Github stars," and that it "committed additional code from our private repos to the Apache 2.0 repo to give the strongest continuity story." Replit "will continue to contribute back to the project as it extends it for enterprise context."

Apache 2.0 lets you use, modify and fork the published code, which is the practical protection here. What the post does not describe is governance: who merges outside contributions, how release cadence is decided, or whether future enterprise features will stay in the open repo. Those are open questions, not criticisms, and they are worth watching in the repository's issues and release notes over the next months.

What is Foundation, and is the benchmark claim solid?

Foundation is Chroma's offering for organizational context that improves as it is used. The company says it "hit SOTA on the BEAM memory benchmark and trended on Ramp insights immediately after release." Both are Chroma's own statements. We could not find an independent evaluation, and benchmark leadership on a memory benchmark is a moving target. Treat it as a vendor claim until a third party reproduces it.

How Foundation will be priced or packaged inside Replit is not stated, and neither is when Replit will ship Chroma-backed memory and retrieval APIs.

Why would Replit want this?

Replit's pitch, in both companies' framing, is a "self-driving company" where agents do a growing share of the work because they have organizational context. An agent platform without durable memory and retrieval forgets between sessions; owning the retrieval layer is a way to fix that and to differentiate from competing agent products.

The agent-memory space is crowded and moving. Related coverage on explainx.ai includes Cognition's Devin dreaming memory spec, Databricks MemCurrent, and Asymptote's cross-harness agent memory. For the retrieval side, see Perplexity's pplx-embed v2 and NVIDIA's Nemotron rerank. Replit has also been in the news over its model supply, as in Replit no longer working with Anthropic, which gives context on why owning infrastructure pieces appeals to it.

What to do this week

If you use Chroma Cloud:

1. Find your date. Open the wind-down timeline Chroma links and write the date down.
2. Export everything. Collections, documents, embeddings, metadata. Keep collection IDs.
3. Pick a destination. Self-hosting open-source ChromaDB is the lowest-friction move because the API stays the same. Alternatives exist, but switching engines means re-testing retrieval.
4. Keep the embedding model fixed. Changing it forces re-embedding and changes results; do that as a separate project.
5. Test before cutover. Run a fixed set of queries on old and new, comparing recall and latency, and confirm backups.
6. Update credentials and endpoints only once tests pass, then cut over reads and writes.

If you use self-hosted ChromaDB:

• Pin your version and read the release notes for the next releases; the extra code moved into the repo may change what the open-source build includes.
• Consider a thin abstraction around your vector store so a future switch is cheaper. Our look at a compositional RAG study from IBM shows how many retrieval choices interact, which is a reason to test any swap rather than assume equivalence.

If you were considering Chroma for a new project: the open-source build is unaffected today, but weigh the uncertainty about governance and roadmap, and prefer a design you can move.

What people are asking

Can I keep using ChromaDB in production after the cloud closes? Yes, if you host it yourself. The open-source database does not depend on Chroma Cloud, per the announcement and the coverage of it. You take on the operations work the managed service did: backups, upgrades, scaling and monitoring. For small and medium collections that is a modest burden; for very large multi-tenant workloads, plan capacity and failure testing before you cut over.

Will the API change? Nothing in the announcement says so. The safest assumption is that the client library and server behave as today, but pin versions and read release notes, since the company says it moved extra code into the open repository and Replit plans to extend the project for enterprise context.

Is this a sign vector databases are being absorbed into agent platforms? It is one data point, not a trend proof. Retrieval and memory are becoming part of the agent stack, and a platform that wants agents to remember an organization's knowledge has a reason to own that layer. Standalone vector stores still serve many teams that are not building on any one agent platform.

Does this affect Replit users right away? The post describes a plan, not a shipped feature. It names no date for Chroma-backed memory or retrieval inside Replit, so nothing changes for Replit users until Replit says more.

How to judge a replacement

If you decide not to self-host, evaluate alternatives on the same few criteria rather than on feature lists:

• Data portability. Can you export documents, vectors and metadata in a standard format, and import them without re-embedding?
• Filter and metadata support. Retrieval quality in agents often depends on metadata filters, not just nearest-neighbor search, so check that your filters behave the same way.
• Latency and cost at your scale. Measure with your real query mix, because published benchmarks use different data and settings.
• Licensing and governance. The Chroma episode is a reminder to prefer permissive licenses and a project that you could fork if needed.
• Exit plan. Whatever you pick, write down how you would leave it, including the export command and how long it takes.

That last point is the lesson of any managed-service shutdown: the cost of switching is set when you design the system, not when the notice arrives.

Open questions

• Exact Chroma Cloud shutdown dates, migration tooling and support options.
• Deal terms and whether Chroma's team stays together inside Replit.
• Foundation pricing, packaging and whether it remains available to non-Replit users.
• Enterprise controls such as data residency and retention for the Replit-hosted context layer.
• How outside contributions to ChromaDB are governed going forward.

Bottom line

Verified from Chroma's own post: a definitive agreement to join Replit, a 60-day cloud wind-down, open-source Chroma continuing under Apache 2.0 with additional code released. Not disclosed: price, exact dates, Foundation packaging. For builders, the actionable part is clear and time-limited: if you are on Chroma Cloud, start migrating now.

Details are accurate as of October 23, 2026 and may change as Chroma and Replit publish more.

Related reading

• Cognition Devin dreaming memory and the open agent-memory spec
• Databricks MemCurrent agent memory
• Asymptote cross-harness agent memory
• Perplexity pplx-embed v2
• NVIDIA Nemotron rerank 8B
• IBM LASAGNA RAG compositionality
• Replit and Anthropic: claimed vs verified
• Primary source: Chroma is joining Replit