Hillock

Hillock is a local neuro-symbolic memory engine for AI agents that need factual recall without a heavy vector database. The public README describes a gradient-free architecture that combines SQLite knowledge triples, Hebbian plasticity, and a 10,000-dimensional hyperdimensional computing reservoir. Instead of sending every query to a model, Hillock first checks whether the question has enough verified local support. Only after a deterministic similarity gate passes does it call a local LLM through Ollama. At review time the repository showed 76 GitHub stars, 7 forks, an AGPL-3.0 license, Python 3.10+ requirements, and a latest visible update date of 2026-08-30. The main idea is useful for local agent builders. Many agent memory systems lean on embeddings, vector databases, and large extraction models. That can work, but it burns memory and can still return plausible answers when the local store does not actually know the fact. Hillock takes a different path. It stores hard subject-predicate-object triples in SQLite, learns associative patterns through Hebbian weights, and uses HDC/VSA matching to decide which facts are eligible for answer generation. If the gate fails, the documented refusal message says it does not have verified information. The README makes several performance and architecture claims that matter for edge users. It targets ingestion and query workflows that can run under 1.2 GB VRAM or even CPU-only, with Ollama used after retrieval rather than throughout the whole pipeline. Version 0.6.0 adds HYDRA token-level MaxSim gating, sub-dimensional projection for faster rejection of weak candidates, and HyperGraph-HDC for multi-hop path reasoning. These details make Hillock more than a demo wrapper around a chatbot; it is an experiment in memory design for machines with limited resources. Hillock is still a technical project. Users should expect to read the README, install Python dependencies, run verification scripts, and understand the tradeoff between deterministic gates and generative answers. The local LLM nuance is important: Hillock is offline and local-first, but answer rendering can call a local Ollama model after the memory system has selected supported facts. That keeps data on the user machine while still using a model for natural-language responses. Pricing is open-source pricing. The code is available under AGPL-3.0, and users pay for their own hardware, local models, and maintenance time. The best fit is an AI agent builder, robotics or edge-AI researcher, or local-first developer who wants a memory layer that can refuse unsupported questions instead of inventing answers. For OpenTools readers, Hillock is worth tracking because it explores a real pain point: compact, inspectable memory for offline agents that need facts, not vibes.