A Relation Is a Rotation On a Knowledge Graph

2026-07-25 · Source: Agus’s Substack · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Knowledge Graphs & Semantic Technologies · Depth: Expert, long

Summary

The `knowlytix` tool implements a geometric memory system where knowledge graph relations are represented as rotations in a 32-dimensional real vector space. This system, demonstrated on a bank's compliance policy, allows AI agents to accurately assess fact plausibility and workflow steps. For instance, an overdraft fee of "35.0" scores a distance of 0.052, while a wrong fee like "45.0" scores 1.527, effectively distinguishing correct from incorrect information. The system handles 40 entities, 8 relations, and 41 triples, ensuring relations are genuine rotations that preserve length and have inverses. Crucially, it supports non-commutative relation composition, essential for complex reasoning chains like escalation paths, and assigns an "admissible radius" to each relation to define its strictness.

Key takeaway

For AI Engineers developing agents for regulated environments, integrating `knowlytix` can address the "prompt and pray" problem by providing a robust geometric memory. This system ensures factual accuracy and workflow compliance by validating claims against policy with calibrated distances and admissible radii. You can move beyond simple string lookups to a memory that flags inconsistencies and provides definitive yes-or-no answers, crucial for agent reliability in sensitive applications like banking compliance.

Key insights

Knowledge graph relations can be modeled as rotations in high-dimensional space, scoring fact plausibility by geodesic distance.

Principles

• Relations as rotations preserve length and have clean inverses.
• High-dimensional rotations enable non-commutative relation composition.
• Admissible radii define relation strictness for fact validation.

Method

`knowlytix` parses policy documents into typed triples, trains a geometry placing entities on a unit sphere, and makes each relation a rotation (rotor) to score fact plausibility via geodesic distance.

In practice

• Use `knowlytix.score_triple` to validate facts.
• Apply `knowlytix.link_predict` for ranking candidate tails.
• Employ `check_holonomy` to detect reasoning inconsistencies.

Topics

• Knowledge Graphs
• Geometric Embeddings
• Relation Learning
• AI Agent Memory
• Compliance Automation
• knowlytix

Code references

• asudjianto-xml/MRM

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• A Relation Is a Rotation — RotatE models knowledge graph relations as rotations in the complex plane, resolving symmetry issues and enabling efficient inference.
• Agentic Knowledge Graphs: why Graphs are becoming the memory layer of AI Agents — AI agents require structured memory and relational reasoning, not just longer prompts, for real-world utility.
• A Relation Is an Operation — Knowledge graphs represent facts as geometric operations, enabling link prediction through spatial transformations.
• LLM Friendly Graphs for AI Memory — From Triples to Facts — Larger, context-rich labels and subgraph representations enhance LLM comprehension of complex facts in knowledge graphs.
• Knowledge Graph Engineering For Agents: Architecting For Self-Improvement — Agent reliability hinges on complete knowledge graphs; early deployment without robust knowledge leads to failure.
• Vector RAG Isn’t Enough — I Built a Context Graph Layer for Multi-Agent Memory — Multi-agent systems need context graphs to combine disparate facts and overcome structural limitations of flat transcripts or vector search.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Agus’s Substack.