After years as a PM in tech, I've worked with all kinds of databases — Oracle, MySQL, Postgres, Redis, Snowflake, and of course MongoDB. Every time you pick a database, you're really picking a way of seeing data.

MongoDB's rise hasn't been about being "the most technically powerful." It's about the fact that from day one, it challenged the most fundamental worldview of the traditional relational database.

Relational databases are built on one assumption: the world can be broken into neat tables, and every column's meaning must be defined in advance. Schema-first — before you write your first line of code, you decide exactly what every field looks like.

That assumption worked beautifully for 40 years. Orders, customers, transactions — that kind of data is naturally structured, and SQL is a perfect query language for it. But underneath that assumption sits another layer: the thing reading and writing the data is an application.

The AI era broke that assumption.

The data an AI Agent needs to handle looks completely different: conversation logs, tool-call traces, intermediate reasoning states, long-term memory across sessions. What all of these have in common — the schema changes every day.

The way AI understands data is, at its core, JSON. An LLM's context, the inputs and outputs of tool calls, the storage of memory — all of it is key-value pair structure.

Flip it around: if an AI needs to read a relational database, it first has to generate a SQL query, send it out, get back a table, then parse that into an object before it can do anything with it. Every step is friction. And LLMs have never been great at writing accurate SQL against complex schemas.

With a document database, the AI reads and writes directly. The shape of the data and the shape of what's in the Agent's "head" speak the same language.

Just some observations from someone who works in tech — not investment advice.