MongoDB, at its MongoDB.local NYC event today, made an aggressive case for using a document database based on the JavaScript Object Notation (JSON) format to more efficiently provide the level of persistent memory that artificial intelligence (AI) agents require.
Company CEO Dev Ittycheria told conference attendees that as databases become the single source of truth for AI agents the way they access data is going to prove crucial. Today, AI agents are loading massive amounts of data in context windows that are costly to process. Going forward, it will soon become feasible for AI agents to process smaller amounts of data while still maintaining the context that AI agents require to provide meaningful outputs.
That approach will make it simpler for developers to build AI agents and applications in a way that reduces the total number of tokens required to process that data. In effect, the database will become the gateway to meaning, notes Ittycheria. In fact, it’s because of this issue that a recent MIT study found nearly all AI projects are failing, he adds. “Memory is so crucial you need to think long and hard about the database that provides these capabilities,” says Ittycheria.
MongoDB is currently working toward enabling more efficient queries using a set of technologies it gained with its acquisition of Voyage.ai earlier this year. Those capabilities include a means to reduce the size of the chunks of data that an AI agent needs to process and a more efficient mechanism for ordering vector embeddings without losing context. The voyage-context-3 model enables full document context to be maintained without having to rely on metadata hacks, summaries provided by large language models or pipeline gymnastics. Additionally, MongoDB is developing a set of granular role-based access controls that in addition to improving governance will also limit the pool of data any AI agent is allowed to access.
As part of the effort, MongoDB is also now moving to embed support for search and vector search across all the iterations of its database to reduce complexity by eliminating the need to integrate and manage external search and vector search platforms.
At the same time, MongoDB has also developed its own set of AI agents for modernizing existing legacy applications. That effort is critical because if organizations are going to add AI capabilities to legacy applications they will need to first turn them into modern applications that make it simpler to expose application programming interfaces (APIs) through which they can invoke large language models (LLMs). Ultimately, MongoDB is moving toward reducing the steep learning curve that developers today encounter when building AI applications.
It may be a while yet before document databases based on JSON become the de facto gateway for memory for agentic AI applications, but the implications are profound as the time, effort and cost of building these applications start to decline. Today the number of AI agents that organizations are willing to deploy in production environments is largely hindered by cost and complexity that MongoDB expects to be able to sharply reduce.
The challenge, as always, is determining what level of investment commitment to make based on the technologies that are available today versus what might soon be possible only a few months later.
