Hillock v0.4: Open-Source Neuro-Symbolic Memory Engine Running Under 1.2GB VRAM
Equivalent-Flan-1590 · reddit · 2026-08-14
A developer released v0.4 of Hillock, an open-source memory engine designed for local LLM agents, aiming to solve the issue of traditional vector databases and LLM extraction processes burning too much context and VRAM.
- Efficient Parsing: Parses documents into plain SQLite SPO triples via a CUDA bi-encoder pipeline (Fastcoref + MiniLM + GLiREL) in 5 seconds.
- Fast CPU Inference: Query gating and coreference resolution run on CPU in <1ms using 10,000-D VSA hypervectors.
- Minimal Footprint: Local LLMs (via Ollama) are called ONLY for final response rendering. The entire engine runs 100% offline in <1.2GB VRAM on a GTX 1070.
The new version also introduces O(1) schema type constraints, auto-direction correction for inverted relations, and regex entity cleaning.
More from coding & agent
- DeepSeek Open-Sources Agent Harness Framework: Everything is a Plugin — AccBalanced · 2026-08-14
- Developer Proposes Using Client-Side AI Agents to Simulate Users for UX Testing — Fowe · 2026-08-14
- MiniMax H3 Combined with Turbo LoRA and SeedVR2 for Local HD Video Generation — cocktailpeanut · 2026-08-14
- Ditching Permanent Infra: A Guide to Entirely Serverless AI Inference — rseroter · 2026-08-14
- TSRX: A TypeScript Extension for Declarative UIs in the Agentic Era — DanWahlin · 2026-08-14
- Cursor Integrates Gemini 3.7 Flash, Shares Internal Eval Results — kalpeshk2011 · 2026-08-14