Richmond Alake, staff developer advocate for AI/ML at MongoDB, recorded in person at the company's London office, argues that MongoDB's JSON-like document model makes it the single database an AI application needs: flexible while teams experiment, lockable in production, and able to hold operational data, metadata, vectors, and even an agent's tools and memory. He frames the coming period as an "era of multi" in which memory management becomes central. The host says MongoDB does not sponsor the episode.
Why documents fit AI. Alake contrasts relational databases' tables, around since the 1980s, with MongoDB's key-value documents. Seeing data represented the same way in the application and data layers, he says, made him fall in love with programming and become a MEAN-stack developer. The host recalls choosing MongoDB for his startups around 2014 for the same flexibility. Alake carries it to AI: requirements shift fast, schema migrations hurt developer speed, which he calls a competitive advantage, and MongoDB lets a schema flex during experimentation and be locked later. With LLMs now emitting JSON through structured output, he argues, the application, the data and the model all speak JSON. The host's example is a Llama model returning a field the OpenAI API does not.
Search and vectors in one place. Atlas, the managed service, runs across AWS, Azure and Google Cloud in multiple regions, because "no developer wants to be a database administrator". The host explains vector spaces via word2vec, then RAG retrieval over a million documents by meaning. Alake says MongoDB first absorbed the search engines customers bolted on, then vector search, so lexical, vector and hybrid search run in one database where some teams had run two or three. For layered explanations he points to his AI-stack post: application, tooling and compute for executives; more layers, starting with programming language, for developers.
Memory and tools. His library Memorizz grew from customers' agent problems. As RAG gave way to agentic RAG in late 2024 and 2025, he says, how memory is structured, stored and retrieved began to govern reliability, an area also pursued by Letta, the company behind MemGPT, and by the HippoRAG paper. The common failure is running out of context window, not memory. Since OpenAI's guidance is to put only 10 to 20 tool schemas in context, his "MongoDB as a toolbox" pattern stores perhaps a thousand tools with embeddings and retrieves two or three before each call.
The multi era. Agents went mainstream this year, so Alake looks to multi-agent architectures, multimodal and multilingual embeddings, multi-step reasoning and several retrieval mechanisms beyond vector search. He predicts voice agents will be widely built but not dominant, and that memory will rise, recalling Andrew Ng opening their first meeting, in April of last year, by asking how he models memory in agents.
Strategy and the long view. For a year he calls "show me the money", Alake bases strategy on pillars and the objective truths within them: for developers, GitHub. His arena framework sets tactics to battle in the market, and his team principles are aggressive, centralized, coordinated, resourceful, efficient and data-driven. In a tangent the host says AGI no longer scares him because human and machine strengths seem complementary. Alake allows a chance of "chaos" rather than collaboration between the two intelligences, but holds that AI will not take human curiosity, and hopes it takes what we now call work.