1 Introduction

TL; DR

• Deterministic protocol and architecture for reasoning agents
• Fixed and taxonomized graph with bounded vector acceleration
• Strong focus on long term memory capable of improving without breaking reliability
• Fast and deterministic online inference on an immutable snapshot
• Governed offline memory consolidation and structural graph evolution followed by publication
• Targets enterprise agents that must remain reliable while learning

2 General problem

Interest in graphs within artificial intelligence systems has grown significantly. This momentum is positive because it marks a late but necessary recognition of their value.

However, this enthusiasm is often accompanied by conceptual confusion. The term graph is frequently used in a generic way without clarifying its scope or its actual role. This ambiguity makes the design of robust and controllable architectures more difficult. Understanding the application context of a graph is therefore a mandatory step before any architectural decision.

3 Graph typology

The term graph covers several distinct realities in artificial intelligence systems. Depending on its role, a graph may be used to represent knowledge, to structure temporary context or to drive an agent’s reasoning and actions.

Understanding these differences is essential because each type of graph addresses a specific objective and operates within a defined scope. Confusing them leads to imprecise and hard to control architectures. Three main categories of graphs can be identified in these systems.

3.1 Knowledge Graph (KG)

The KG is a structured knowledge graph including entities and relations. It serves as the fundamental knowledge base modeling relatively stable and shareable information. To ensure this data is usable by an agent, SHACL is often applied to validate its compliance and structure. It is used to organize domain knowledge so that the agent can access verified and error free facts.

3.2 Context Graph (CG)

In addition to stable knowledge, the context graph is built dynamically for a specific query or task. It selects and organizes only the information that is relevant in order to provide an adapted context. Its goal is not completeness but immediate relevance.

3.3 Reasoning Graph (RG)

The RG is a control graph for reasoning retrieval and agent actions. It does not only contain data but structures the decision process by integrating environment specific business reasoning. It relies on the rigorous structure validated in the KG to define logical steps and control how information is used to trigger concrete actions. It plays an active role in execution rather than being limited to representation.

4 Retrieval and semantics

Building systems capable of handling complex information requires relying on standards that allow data to be structured and queried precisely. This precision is first achieved through rigorous data models.

4.1 RDF and SPARQL

The RDF data model is based on a triple structure composed of subject, predicate and object. This organization makes it possible to explicitly connect data elements. SPARQL then acts as the query language used to interrogate these triples and extract information in a structured way.

4.2 Ontology taxonomy and OWL

Alongside query languages, an ontology defines the concepts of a domain and the possible relationships between them, while a taxonomy organizes these concepts hierarchically. OWL is a language used to formalize these semantic structures. In practice, adapting OWL to a specific domain often requires many iterations, which can limit its operational effectiveness.

5 Limits of traditional approaches

Traditional semantic graphs are mostly centered on domain or enterprise data. They excel at structuring static knowledge but show limitations when it comes to integrating dynamic contexts, user interactions or agent actions. This narrow focus makes it harder to build truly adaptive systems.

6 Adaptive Reasoning Graph (ARG)

The Adaptive Reasoning Graph proposes a broader and structured approach to graph based reasoning in order to address these limitations. It enables the simultaneous exploitation of:

• Taxonomy
• Ontology
• Reasoning
• Context
• Memory

Unlike traditional graphs, ARG can integrate user memory in graph form, business or organizational knowledge and information sourced from the web. All of this is brought together within a single coherent and harmonized environment.

7 Graph memory

Memory is a central component of agent systems that must reason and act reliably over time. In a graph based architecture, memory becomes an exploitable and traversable structure governed by explicit rules and expressed through two complementary forms.

7.1 Short-term memory

Short-term memory corresponds to the language model context window. To extend the effective context, non graph approaches exist such as recursive language models. However, it is generally preferable to filter and rerank information first, then build a refined chunk based context in order to optimize relevance without unnecessarily increasing system complexity.

7.2 Long-term memory

In contrast, long-term memory aims to preserve and evolve information over time. Solutions such as Mem0 or MemGPT help manage this memory, but they are not graph native. This results in the loss of explicit taxonomy and ontology structures. ARG provides an alternative that preserves these properties while ensuring an evolving memory.

8 Action and agent orchestration

Modern agents are no longer limited to producing responses because they also execute tasks. This execution traditionally relies on several distinct components that ensure system reliability:

• The PLANNER component defines the action plan.
• The CHECKER or VALIDATOR component verifies correctness.
• The ORCHESTRATOR component coordinates execution.

In an ARG based architecture, the language model operates by explicitly respecting graph rules. This approach is comparable to ReAct, but with a reasoning graph that makes behavior more deterministic and more controllable. Control remains at the architectural level rather than being fully delegated to the model.

9 Architectural choice

Two main strategies emerge when designing agent systems. The first is to build a multi agent system composed of many distinct environments. The second relies on a single coherent, homogeneous and harmonized environment capable of covering most use cases. The Adaptive Reasoning Graph follows this second approach.

10 Conclusion

Using graphs in agent systems requires a clear understanding of their role and scope. By distinguishing knowledge graphs, context graphs and reasoning graphs, it becomes possible to design more reliable and controllable architectures. The Adaptive Reasoning Graph provides a unified framework that integrates context, memory and action within a structured environment, enabling agents to learn and evolve while maintaining a high level of reliability.