AI Agent with Model Selection, Tool Synthesis, and Proactive Knowledge Base Management

Abstract

This paper presents a concept and pilot implementation of an adaptive AI agent that extends conventional large language model (LLM) systems with planning, tool usage, long-term memory, and continuous self-improvement. Unlike static agent architectures, the proposed solution operates on three levels of adaptability: operational adaptation during task execution, task-specific adaptation through autonomous creation of new tools, and long-term adaptation based on accumulated experience and knowledge restructuring. The architecture integrates a user interface, an orchestrator, multiple external language models, a dynamic tool and skill layer, a knowledge base, and an evaluation layer supporting interoperability. A key contribution is the integration of multi-model routing, proactive memory refinement, user feedback, and autonomous tool synthesis within a single framework. Initial experiments using a local Ollama model server demonstrated the agent's ability to generate specialized tools for accessing external resources such as arXiv, reuse previously acquired experience, and improve efficiency in subsequent tasks. Combining memory-driven learning, reflection, and tool generation could increase adaptability when solving heterogeneous tasks. The proposed platform also creates opportunities for future research in long-term evaluation, knowledge representation, and graph-based memory structures.