Valori: A Deterministic Memory Substrate for AI Systems

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: Gudur, Varshith
Natura: Recurso digital
Lingua:inglese
Pubblicazione: Zenodo 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866901398245867520
author Gudur, Varshith
author_facet Gudur, Varshith
contents <p>Modern AI systems rely on vector embeddings stored and searched using floating-point arith-metic. 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 re-places 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 deter-ministic 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.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18022709
institution Zenodo
language eng
publishDate 2025
publisher Zenodo
record_format zenodo
spellingShingle Valori: A Deterministic Memory Substrate for AI Systems
Gudur, Varshith
Deterministic AI
AI Memory
Vector Databases
Fixed-Point Arithmetic
Reproducible Computation
Approximate Nearest Neighbor Search
State Machines
Auditability
Systems for AI
<p>Modern AI systems rely on vector embeddings stored and searched using floating-point arith-metic. 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 re-places 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 deter-ministic 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.</p>
title Valori: A Deterministic Memory Substrate for AI Systems
topic Deterministic AI
AI Memory
Vector Databases
Fixed-Point Arithmetic
Reproducible Computation
Approximate Nearest Neighbor Search
State Machines
Auditability
Systems for AI
url https://doi.org/10.5281/zenodo.18022709