Hillock

A lightweight, 100% local neuro-symbolic memory engine built for edge hardware.

A local memory engine that answers from a knowledge graph, not a vector database, so there is no drift and no approximate matches standing in for facts.

Ingests a document in ~5 seconds and runs the entire pipeline in under 1.2 GB VRAM (or CPU-only), instead of the 5-16 GB+ and long generation waits an LLM-based extraction pipeline needs.

A hard, deterministic similarity gate blocks unanswerable questions before they ever reach the LLM, so it refuses honestly instead of generating a plausible-sounding guess.

100% offline: SQLite for facts, Hebbian weights for associative recall, and a 10,000-D hypervector space for sub-millisecond context matching. Ollama is only called once a question has already passed the gate.

Traditional local RAG is surprisingly heavy. Running dense vector databases and using 8B+ generative LLMs just to parse documents and maintain long-term memory burns VRAM, chokes mid-range GPUs, and still hallucinates when asked about things it doesn't know.

Hillock was built to solve this. It replaces bloated vector databases and token-hungry extraction passes with a lightweight, three-tier architecture combining relational knowledge graphs, Hebbian synaptic memory, and 10,000-dimensional Vector Symbolic Architectures (VSA/HDC).