Beyond Vector Databases

One of the pillars of the LLM stack is a vector database. Vector dbs use what’s called vector embeddings to encode unstructured data, so that you can take documents, images, video, and other data and query over them. They’ve become indispensable for the Retrieval Augmented Generation (RAG) framework that has become a staple of many chatbot applications.

Yet vector databases have their limitations. What to do when unstructured data contains facts that we want to extract for exact operations? What if a piece of unstructured data contains the same entities and relations referenced in a row of structured data? Similarity comparisons leave us helpless. The gulf between structured and unstructured data left open by vector databases means we lack unified data models and seamless backends.

The result of this arbitrary division is that if you want an LLM-centered app to work with structured data, you need to create a complicated backend to merge and integrate various data silos, or trust in wonky, failure-prone SQL generation or proxy techniques. This lack of a unified seamless backend also means we simply can't engage in automated analytics at true scale. What's needed is a continuous, fully integrated data layer where these arbitrary distinctions between data types no longer get in the way.

To address this most in the industry are creating various SQL wrappers, but this is just a half-measure. They still require separate systems to be glued together. Such fragmentation means there is no way to create a universal data model.

The math behind vectors, the core concept of vector databases, is much more powerful than existing vector databases suggest. It's been known for years that vectors can serve as richer representational models of the world, and that they can be used to represent the relationships between entities and carry discrete symbols. While this knowledge has remained in the realm of academic research, Semantic Reach is going above and beyond by tapping into its commercial potential like never before. It turns out you can actually do a lot of processing and analysis directly the vector space, which means that real time analytics and intelligent fact discovery and inference can be performed. This isn't a marginal improvement on existing technology. This is a whole new class of capability.

So that's how we’re creating what we call Deep Memory. The key insight is that if you’re building an AI-centric app, what matters is not how the data is stored, but how it is represented to the model. The data model should be designed in such a way that it organizes information that enhances the model’s ability to reason about the world. That’s why one of our taglines is “From Database to Databrain.” We aren’t just storing data, we’re organizing facts and relationships usings applied cognitive science research and machine learning techniques. So when the model goes to access our backend, it’s not just retrieving isolated data points, it’s returning the meaning in the data. Analytics and storage should be unified so that AI models tap into structured knowledge and leave out the hallucination-prone guesswork. This intelligent data layer distinguishes us from RAG, which is just a retrieval framework.

The data layer should be smart, applying machine learning to continuously adjust and adapt to the evolving connections latent in the data. The net result is models that are faster, token-efficient, more accurate, and agentic systems that can coordinate and remain on task much more effectively via a single source of truth and context. Our memory solution is the first to be truly AI and vector native.

Because of the math we’re using, our memory solution works with all kinds of data, structured and unstructured, and comes prebuilt with provable audit trails that the system can use to prove to itself that the recovered value is what it says it is. Sameness is just the limiting case of similarity: our vector architecture goes the extra mile to be able to retrieve, compare, and compute both by comparison and identity. This unlocks a whole slew of new capabilities. It gives us AI backends that can support factual reasoning and traceable assertions.

You can’t have true intelligence without memory. Memory is a basic cognitive primitive. LLMs lack this concept, and therefore can't do continual learning. RAG systems also fail to emulate cognitively plausible memory. LLMs need a memory that differs from the rigid type used in computers, and that is more powerful than the surface-level text prompt solutions currently in use. True cognitive memory adapts and readjusts associative connections to enhance predictions and inference. A proper memory backend should build off the facts and patterns that have already been intelligently processed in the data layer and provide the skeleton of reasoning. This is the path to hallucination-free, accurate, powerful, and reliable AI performance.