Samsung's Acquisition of Oxford Semantic: RDFox and Neuro-Symbolic AI

Should we trust AI with our schedules and health records?

Samsung’s acquisition of Oxford Semantic (July 2024) secured the core engine for personal knowledge graphs within devices. It is widely interpreted as a strategic move to address the limitations of Generative AI by incorporating logical reasoning through Neuro-Symbolic AI.

What is the goal behind Samsung’s acquisition of Oxford Semantic?

On July 16, 2024, the University of Oxford announced that Samsung Electronics had acquired its spin-out company, Oxford Semantic Technologies (OST). On July 18 (KST), Samsung Electronics stated via its Newsroom that it had been collaborating with OST since 2018 and had previously invested in the company through Samsung Ventures. The acquisition provides a core engine for a personal knowledge graph that aggregates information and context scattered across services and apps, enabling hyper-personalized experiences while keeping sensitive personal data on-device. The scope of application is intended to extend beyond smartphones to TVs and home appliances.

This is where the official announcements end. The interpretation that “this is a move to reduce hallucinations” is my own, and this post distinguishes between the two. The three technical pillars discussed later are the knowledge graph (the brain), semantic communication (the nervous system), and semantic segmentation (the eyes).

How does Neuro-Symbolic AI reduce hallucinations in Generative AI?

Large Language Models (LLMs) like ChatGPT are probabilistic models that calculate the probability of the next word based on vast amounts of training data. The reason their sentences are impressively fluent, as well as why they struggle to distinguish truth from falsehood—leading to hallucinations—stems from this same structure. If an AI is to evolve from a conversational partner into a personal assistant that manages our schedules, this weakness must first be addressed.

Neuro

This is the knowledge graph and reasoning engine of OST, founded in 2017 by Oxford researchers (Ian Horrocks, Boris Motik, and Bernardo Cuenca Grau). While relational databases like Excel struggle with tables, a knowledge graph stores connections such as “Chul-soo is a friend of Young-hee, and Young-hee likes the pasta place at Gangnam Station” as triples: (Subject)-[Relationship]-(Object).

User → Located at → Gangnam Station → Feature → Traffic Congestion

As data grows, the context gained from these connections becomes richer. There are other graph databases like Neo4j on the market, so Samsung’s choice of RDFox is worth examining. The RDFox product page highlights three features:

• In-memory: Designed to run on RAM rather than disk. While the company claims loading and query times can be 10–1000 times faster, this is a claim from the developer without independent benchmarks. I am correcting the “100–1,000 times faster” statement from the first edition as it was an unconditional assertion.
• Incremental Reasoning: When data changes, it dynamically executes rules to re-process only the affected information. This structure is advantageous for resource-constrained environments as it avoids recalculating everything from scratch.
• Datalog Rules: Logical rules can be inserted directly into the engine. An example would be, “If a user is at a subway station and it is 8 AM on a weekday, they are commuting,” though this is also a hypothetical example for conceptual explanation.

What is confirmed as the reason for Samsung’s choice is a statement from co-founder Ian Horrocks in an interview, where he mentioned that Samsung was interested because RDFox runs directly on mobile phones with low computational overhead. There is no basis to claim this is the “only engine capable of this,” as comparative data with other engines has not been made public.

What will the RDFox-based personal knowledge graph change in the Galaxy S25?

In a blog post on January 22, 2025, OST stated that RDFox is the underlying technology for Samsung’s Personal Data Engine and is included in the Galaxy S25 series. According to Computer Weekly, Professor Horrocks explained that this engine stores the user’s experience in their personal knowledge graph. This shifts the direction toward an AI of one’s own that focuses on a Personal Knowledge Graph (PKG), rather than an “AI for everyone” that aggregates the world’s knowledge on a server. The flow of on-device processing was also summarized in my On-Device AI post.

Regarding privacy, I am correcting the “perfectly protected” claim from the first edition. While Samsung explains that the personal data engine processes data within the device and is protected by the Knox Vault security hardware, features that use cloud AI follow a hybrid approach, accessing data only when necessary and deleting it immediately after the request is processed. The goal is to keep it securely on-device, but I have not been able to confirm independently verified results.

Sometimes the Galaxy S25’s “Now Brief” is introduced as a result of this acquisition, but there is no confirmed connection. Neither the OST blog nor Computer Weekly directly linked Now Brief to RDFox, and Samsung’s explanation is limited to the feature being a personalized daily summary. The scenario where the AI infers, “It’s raining? Then traffic will be heavy,” and notifies you to “leave 20 minutes early to not be late for your 9 AM meeting” should be read only as an example of what is possible when a knowledge graph connects schedules, weather, and traffic.

My initial forecast that a “lock-in effect” would occur—where users find it hard to switch from Galaxy because of the accumulated data graph—was my own estimation. There is no official material to support this.

What do the neural network pillars of Neuro-Symbolic AI, semantic communication, and DeepJSCC promise?

The roots of semantic communication lie in communication theory. The three levels of communication problems are commonly attributed to Claude Shannon, but they were actually a framework presented by Warren Weaver while explaining Shannon’s theory. Shannon himself drew a line, stating that the semantic aspect was irrelevant to engineering problems. Weaver’s three levels are:

• Level A (Technical problem): How accurately can symbols be transmitted?
• Level B (Semantic problem): How precisely do the transmitted symbols convey the intended meaning?
• Level C (Effectiveness problem): How effectively does the received meaning influence behavior in the desired way?

4G and 5G could be considered a Level A competition to transmit “faster and more.” Research into semantic communication is an attempt to elevate communication from Level A to Level B. Since I could not find any official material stating this as Samsung’s goal, I am treating it only as an industry research direction in this post.

DeepJSCC (Deep Joint Source-Channel Coding) is a representative technology in this flow. In a 2019 paper, Bourtsoulatze, Kurka, and Gündüz proposed a neural network that converts image pixels directly into channel signals without compression codes or error correction codes. It showed “graceful degradation,” where image quality drops gradually even as channel conditions worsen, contrasting with digital methods where signals collapse suddenly once they fall below a threshold. I am removing the “data reduction by about 1/1000” and “transmitting only eye/nose/mouth vectors” explanations from the first edition as they are not in the paper. The bandwidth ratios tested in the paper range roughly from 0.04 to 1.

Ideas such as sending only text and sketch information, with a generative model like Stable Diffusion reconstructing the image on the receiving end, are also discussed as research directions. However, since I could not verify figures like compression rates or image quality in public materials, I have not included them. The expectation that it could be used in remote or satellite environments remains just an expectation for now.

What does Qualcomm’s semantic segmentation, the eyes of Neuro-Symbolic AI, do?

Semantic segmentation is a technology that breaks video into frame-by-frame segments to recognize sky, skin, hair, and buildings separately. In its Snapdragon 8 Gen 3 specifications, Qualcomm introduced a Cognitive ISP that performs real-time semantic segmentation, dividing photos and videos into up to 12 layers and optimizing each. For example: “This area is the sky, so let’s make it bluer,” or “This area is skin, so let’s process the texture to be smoother.” The difference from simple filter correction is that the processing changes according to the meaning of the area.

One correction: the first edition described this as a feature of the Snapdragon chipset used in Galaxy, but the Galaxy S25 series uses the Snapdragon 8 Elite (for Galaxy) for all models, and Cognitive ISP is a term written in the 8 Gen 3 specifications. I have not confirmed the basis for directly linking it to S25 camera features. The concept that segmentation results (“the user is currently on a mountain”) are passed to the knowledge graph to enhance context is also a scenario I imagined, and I have not found any publicly implemented examples.

Where does Samsung’s acquisition of RDFox stand in the era of Agent AI?

The first edition stated, “We will witness Agent AI after 2025.” As of October 2026, this is no longer a forecast but a verification stage. Gartner released a report in June 2025 stating that by 2028, at least 15% of daily work decisions will be autonomously handled by Agent AI, while over 40% of Agent AI projects will be canceled by the end of 2027. It is a time of both high expectations and trial and error.

In this flow, the significance of Samsung’s acquisition lies in securing a foundation for accumulating personal data into an on-device knowledge graph. My judgment is that future AI competition will shift from “who makes the bigger model” to “who understands the user more deeply and logically (semantically).” A comparison post on this same site regarding Semantic AI and Palantir Ontologies points to the same logic in the enterprise ontology sector.

The table below divides the contents of this post into confirmed facts and interpretations.

Claims regarding Samsung’s acquisition of Oxford Semantic Technologies	Category	Basis
Acquisition announced in July 2024, secured personal knowledge graph engine	Confirmed	Samsung Newsroom, University of Oxford
RDFox is the foundation for the Galaxy S25 personal data engine	Confirmed (Developer announcement)	OST Blog
In-memory, Incremental Reasoning, Datalog	Confirmed (Developer explanation)	RDFox Product Introduction
Neuro-Symbolic foundation to reduce hallucinations	Interpretation	No public figures
Link between Now Brief and RDFox, lock-in effect	Unconfirmed	No official material
Integration of semantic communication/segmentation with knowledge graph	Conceptual	No public implementation examples