IDE-Cavernícola is an experimental artificial intelligence research project exploring an alternative to the current paradigm based exclusively on LLMs. Its approach, known as Inventor Driven Cognition (IDC), stems from a simple idea: intelligence is not about remembering increasing amounts of information, but about extracting useful principles, validating them against reality, and retaining only what continues to work. While a traditional LLM follows a cycle of prompt → response → forgetting, IDC proposes a cycle of goal → hypothesis → validation → causal learning → persistent knowledge. The system utilizes episodic memory to record experiences, causal memory to store validated rules, and a negative memory termed Causal Trash to avoid repeating known mistakes. In current prototypes, IDC integrates with Gemini, which serves as a creative hypothesis generator, while IDC executes, validates, and decides what knowledge is worth preserving. If a strategy fails, it is recorded as a forbidden pattern; if it succeeds, it is consolidated into a reusable rule. The goal is for the agent to learn once and reuse that knowledge indefinitely without needing to reprocess all the original information.
Currently, the project is evolving toward a repository-based cognition model, where the source code itself becomes a source of experiences. To achieve this, it integrates tools such as ast-grep, structural analysis for Python, JavaScript, and TypeScript, episodic storage in SQLite, automated detection of technical debt, architectural analysis, CI/CD metrics, and causal hypothesis generation derived from the actual evolution of the repository. IDC seeks to answer a fundamental research question: can an intelligent system become more efficient by storing validated principles instead of accumulating ever-growing amounts of information? The core hypothesis is that intelligence increases when an entity can discover patterns, test them against reality, discard the noise, and retain only the rules that remain useful. In other words: intelligence is not about remembering more, but about forgetting better. Current active areas of development include the causal memory engine, the negative memory system, repository cognition, autonomous optimization, purpose-driven reasoning, and Gemini integration. The project is open to collaboration, discussion, and experimentation by developers, researchers, and AI enthusiasts interested in exploring more efficient, explainable, and persistent cognitive models.
IDE-Cavernícola
Save the pattern. Forget the noise.