7 open roles (see below)
Member of the Technical Staff - Distributed Systems
Member of the Technical Staff - Product Engineering
Member of the Technical Staff - Cloud Infrastructure
Member of the Technical Staff - Database Systems
Member of the Technical Staff - Open Source Stewardship
Member of the Technical Staff - Applied Research
Director of Special Operations - Community
<aside> <img src="/icons/rocket_gray.svg" alt="/icons/rocket_gray.svg" width="40px" /> Chroma’s mission is to build robust and reliable programmable memory for AI.
The emergence of powerful general purpose large language models into the mainstream has made it possible to build practical applications with AI for the first time. Much has been said about these model’s potential and capabilities, but at minimum they represent a completely new way to develop software. The scale of this shift is comparable to the emergence of the web, but the effects will be far greater.
We are in an age of incredible experimentation.
Each new wave of computing since the mainframe - PC, the web, mobile, and cloud - has resulted in the creation of a new software stack. The application logic layer in this new stack is handled by the AI itself, programmed in natural language with total flexibility. But logic without memory is only half of the story.
The memory, storage, and state layer requires a totally new approach - programmable memory for AI. Beyond storing and retrieving information, the programmable memory layer defines the information available to the AI itself. It’s the key to making the AI robust, controllable, interpretable, and safe.
<aside> <img src="/icons/science_gray.svg" alt="/icons/science_gray.svg" width="40px" /> Chroma builds the stack for programmable memory for AI.
At the core of Chroma’s programmable memory engine is our vector database.
In order to match the flexibility of the AI logic layer, the programmable memory layer must store and retrieve data based on its semantic meaning, rather than its structure and content. Today this is accomplished through embeddings-based vector similarity search.
Vectors represent points in high-dimensional space, commonly called latent space. These points represent a map of meaning - where the geometry of the data itself means something. Points and measuring the distance between points (nearest neighbor search), is only the beginning. Vectors have more in common with maps than tables.
Powerful, production, AI memory requires much more than just vector similarity search. It would be like handing a developer a database storage engine, without the rest of the database. Just because a result is the most similar, doesn’t mean it’s actually relevant. LLMs get easily confused by irrelevant context, and this is made worse in the case of long context windows.