Valori: A Deterministic Memory Substrate for AI Systems
Varshith Gudur
2026-01-01
Summary
This paper focuses on a problem with how modern AI systems store and retrieve information, specifically that the results aren't always the same even when everything else is. They introduce a new system called Valori that aims to make AI memory completely predictable and consistent across different computers.
What's the problem?
Current AI systems use a method of storing information called vector embeddings, which rely on complex numbers and calculations. Because of the way computers handle these calculations, even if you use the exact same AI model, the same input data, and the same code, you can get slightly different results on different types of computers, like an Intel-based PC versus an Apple computer. This inconsistency makes it hard to trust the AI, especially in situations where you need to be able to verify its decisions or keep a reliable record of what it did, like in regulated industries.
What's the solution?
The researchers created Valori, which replaces the complex number calculations with simpler, fixed-point calculations. Think of it like switching from using decimals to only using whole numbers and fractions. Valori also treats the AI's memory as a predictable 'state machine,' meaning each step is clearly defined and always produces the same outcome. This ensures that Valori always generates the exact same memory states and search results, no matter what computer it's running on.
Why it matters?
This work is important because it addresses a fundamental issue of trust and reliability in AI. If AI systems aren't deterministic – meaning they don't always give the same answer to the same question – it's difficult to verify their behavior and ensure they're safe and fair. Valori provides a way to build AI systems with predictable memory, which is a crucial step towards creating trustworthy AI, especially in fields where accuracy and accountability are essential.
Abstract
Modern AI systems rely on vector embeddings stored and searched using floating-point arithmetic. While effective for approximate similarity search, this design introduces fundamental non-determinism: identical models, inputs, and code can produce different memory states and retrieval results across hardware architectures (e.g., x86 vs. ARM). This prevents replayability and safe deployment, leading to silent data divergence that prevents post-hoc verification and compromises audit trails in regulated sectors. We present Valori, a deterministic AI memory substrate that replaces floating-point memory operations with fixed-point arithmetic (Q16.16) and models memory as a replayable state machine. Valori guarantees bit-identical memory states, snapshots, and search results across platforms. We demonstrate that non-determinism arises before indexing or retrieval and show how Valori enforces determinism at the memory boundary. Our results suggest that deterministic memory is a necessary primitive for trustworthy AI systems. The reference implementation is open-source and available at https://github.com/varshith-Git/Valori-Kernel (archived at https://zenodo.org/records/18022660).