Dynabolic-RE: A Pure C++ Implemented Chain Graph Reasoning Engine, Completely Free from the Constraints of Matrix Operations

Dynabolic-RE is a graph reasoning engine entirely built on C++ standard libraries. It achieves multi-step reasoning by replacing matrix multiplication with a native graph structure, supporting forward/backward chaining reasoning, Bayesian probabilistic reasoning, temporal trend analysis, and a truth maintenance system, providing verifiable symbolic reasoning capabilities for LLMs. The project was developed and open-sourced by Noverisp3. The core idea is that traditional matrix frameworks are not suitable for graph reasoning, as tools become an obstacle. Therefore, a native graph engine specifically designed for reasoning is built using pure C++ standard libraries.

Modern AI frameworks like PyTorch, TensorFlow, and NumPy are optimized for matrix operations, but graph reasoning tasks are inherently different from matrix operations—implementing graph reasoning with a matrix framework is like using a hammer for surgery; the tool itself becomes an obstacle. Thus, Dynabolic-RE was born, completely abandoning matrix multiplication. It is built using C++17 standard libraries such as <vector>, <map>, <thread>, and <memory>, without relying on any external libraries. It was developed by Noverisp3 and open-sourced on GitHub.

Core Architecture: The graph structure includes GraphNode (7 types: ConceptNode, FactNode, RuleNode, QueryNode, InferenceNode, MemoryNode, ControlNode) and GraphLink (7 relationships: Causal, Implies, Supports, Contradicts, Sequential, Hierarchical, Associative). Link operations are thread-safe.

Chain-of-Links Mechanism: 1. Path tracing (DFS from query nodes to avoid loops); 2. Signal propagation (applying rules according to link types); 3. Causal chain analysis (tracing the basis of conclusions), making reasoning transparent and interpretable.

Basic Logical Operations: Supports complete classical logic including AND, OR, NOT, and IMPLIES.

Non-monotonic Reasoning and Truth Maintenance: Implements a Truth Maintenance System (TMS), including dependency tracking (recording facts and rules that conclusions depend on), recursive revocation (cascading revocation of dependent conclusions), forward chaining derivation (inferring new conclusions from known facts), and contradiction detection (identifying logical conflicts), distinguishing it from simple rule engines.

Bayesian Probability Processing: Integrates a Bayesian processor, supporting confidence calculation (0.0-1.0), posterior probability update, mutual information calculation, and entropy analysis; rule triggering requires the antecedent confidence to exceed a threshold (default 0.7).

Temporal Trend Analysis: Supports timestamp recording of fact validity time, trend detection (attribute change trends), and process understanding (dynamic state reasoning such as "temperature is rising"), suitable for monitoring and diagnostic applications.

Performance Optimization: Sparse graph representation (space complexity O(V+E)), smart pointer memory management, no weight matrices, attribute compression; multi-threaded architecture (producer-consumer model, task queue + worker threads, atomic performance metrics). Node operations are O(1), path finding is O(V+E), multi-threading provides near-linear improvement, with a throughput of approximately 1.25 million nodes/second and 666,000 links/second.

LLM Integration: Provides a Python orchestrator. The workflow is: LLM converts natural language into structured JSON → C++ engine performs reasoning → LLM converts back to natural language. It supports backends like Ollama and OpenAI API, dividing labor to leverage the language capabilities of LLMs and the symbolic reasoning advantages of the engine.

Advantages: Fully interpretable (clear derivation paths), deterministic behavior (consistent output for the same input), editable knowledge (explicit addition, deletion, modification), zero external dependencies (only C++17 standard libraries).

Applicable Scenarios: Expert systems requiring strict logical consistency, decision support systems with high compliance requirements, audit scenarios needing reasoning explanations, resource-constrained edge computing, and hybrid reasoning systems combined with LLMs.

Dynabolic-RE completely breaks free from the constraints of matrix operations through a native graph structure, achieving efficient, transparent, and interpretable symbolic reasoning. Its professional functions such as truth maintenance, Bayesian processing, and temporal analysis are powerful tools for building complex reasoning systems. The combination with LLMs shows broad prospects for the fusion of symbolic AI and connectionist AI, representing an important direction of AI reasoning technology—explainable and verifiable symbolic reasoning.