Complex Knowledge Curation using Agentic Ontological Notebook Memory
Gully Burns (Unaffiliated), Paul Groth (University of Amsterdam)
Architectural Patterns & Composition
Summary
A personal AI research assistant that uses explicit ontological commitments stored in a TypeDB knowledge graph as agent memory, demonstrated for job-hunting and disease-mechanism understanding.
Description
In the same way that every software system uses a data model (as database schema, type system, or API contracts), large language model (LLM)-powered agents makes domain-specific semantic distinctions in prompts, tool definitions, data sources, etc. These ontological commitments are implicit, hard to test, and difficult to refine systematically. Furthermore, the way that agents preserve context across sessions (agent memory) usually make few, if any, domain-specific ontological commitments and rely only on the inherent intelligence of LLMs to make those distinctions. We argue that making these commitments explicit improves agent performance for domain-specific work. To that end, we present Skillful-Alhazen as a personal AI research assistant that uses skills to define domain-specific instructions, tools (coded as scripts), and a new kind of 'ontological notebook' as memory. Users interact with the assistant in plain language, and the assistant applies the semantic definitions from its memory schema to execute work. It retains structured context based on data in the explicitly defined domain model it can access, query, and modify. Skillful-Alhazen uses TypeDB to implement the ontological notebook. This is a recently-commercialized, closed-world, declarative, logic-based persistent knowledge graph with several desirable properties in this application: including type-safety, strong inference, easy refactoring, and scale. We demonstrate the use of Skillful-Alhazen in two domains: job-hunting and disease-mechanism-understanding, each based on separate skill definitions. Our tool is open-source available from https://github.com/sciknow-io/skillful-alhazen
