A developer released Hillock, a lightweight local AI memory system using SQLite and brain-inspired math, as an early-stage experiment to simplify offline chatbot memory without heavy vector databases.

What happened: A developer published Hillock, a personal project that combines SQLite fact storage, Hebbian plasticity (dynamic connection strengthening between entities), and hyperdimensional computing (a 10,000-dimensional vector approach) to manage memory for offline AI chatbots. The system achieved 30.0% retrieval accuracy and 30.0% gate accuracy on a 30-sentence benchmark using a tiny Qwen 1.5B model, with extraction precision at 10.6% and recall at 22.7%.

Why it matters: Standard vector databases felt too heavy and complex for running a simple, offline chatbot on a personal computer. Hillock attempts to prove that brain-inspired math can make local AI memory more practical, though the creator emphasizes this is a work-in-progress with rough edges and bugs—not a finished product.

What to watch: The low benchmark scores (10.6% extraction precision, 22.7% recall) reflect the difficulty of parsing dense, multi-subject sentences with a 1.5B parameter model. The creator provides a GitHub repository and setup instructions, inviting interested users to test the prototype locally with a Python virtual environment and Ollama.