LangChain and MongoDB Announce Partnership: An AI Agent Stack Running on a Database You Already Trust

LangChain and MongoDB have joined forces to bring together everything needed to operate agents—retrieval, persistent memory, access to operational data, observability, and deployment—all based on MongoDB Atlas.

The core message is that running agents on top of an operational database that many enterprises already trust can reduce the complexity that comes from cobbling together a separate vector DB, state store, and analytics API. MongoDB emphasized that more than 65,000 customers run mission-critical applications on Atlas, and presented a strategy of layering an agent backend on top of that.

The scope of integration spans LangSmith, LangGraph, and LangChain as a whole. As a result, Atlas now provides the following capabilities on a single platform.

• RAG based on Atlas Vector Search: retrievers ready to use directly from Python and JavaScript SDKs, with support for semantic search, hybrid search (BM25 + vector), GraphRAG, and pre-filtered queries
• MongoDB Checkpointer: stores an agent's conversation state and execution state in MongoDB, supporting crash recovery, human-in-the-loop, time-travel debugging, and fault-tolerant execution
• Text-to-MQL: takes natural language questions and explores collections, checks schema, and generates, validates, and executes MQL aggregation pipelines
• LangSmith observability: end-to-end tracing covering retrieval calls, tool invocations, routing decisions, and even checkpoint writes

In particular, the checkpointer changes a cost structure that stands out in production environments. In the default architecture, each agent needs a separate Postgres instance to handle high-frequency checkpoint writes, but using the MongoDB checkpointer allows state from multiple deployments to be consolidated into a single shared cluster, simplifying the structure so that N databases become 2.

As a use case, security company Kai Security was introduced. They stated that they were able to add persistent state, pause-and-resume, crash recovery, and audit trails in a single day—capabilities that had previously been blocked by the burden of having to operate a separate data layer.

In summary, this partnership is an attempt to bundle vector search + state persistence + natural language data querying + observability into a single stack, reducing the infrastructure assembly problems most commonly encountered when moving from prototype to production. It also emphasizes offering both open source and managed deployment together, so it isn't locked into any particular cloud or LLM vendor.