← sources

After 6 months of daily building, I just released Hillock v1.0 stable: a local neuro-symbolic memory engine that runs under 1.2GB VRAM with zero hallucinations

source: reddit (reddit:selfhosted)published 2026-10-28 09:52ingested 2026-10-29 02:39

retained text

If you self-host local AI tools, you probably know how annoying it is to set up a memory layer. Most modern setups expect you to spin up multiple heavy Docker containers running dedicated vector databases that hog gigabytes of RAM just to store a couple of text documents. I spent the last six months building Hillock to replace that entire stack with a single, ultra-fast SQLite database and pure CPU math, and v1.0.0 stable is officially out today.



Hillock is designed to be completely self-contained. It stores all facts, relational links, and associative synaptic weights directly in a standard SQLite file using Write-Ahead Logging for fast concurrent reads and writes. To prevent the AI from making things up, it compresses query contexts into hyperdimensional bit vectors and runs hardware-level popcount checks on your CPU. The entire memory engine stays under 1.2GB VRAM and can run purely on laptop CPUs with zero GPU required.



It exposes a standard OpenAI compatible API at port 8000, which means you can plug it directly into your existing self-hosted interfaces like AnythingLLM or custom chat bots without changing your code. It has built-in streaming, local entity inspection commands, and zero external telemetry or cloud tracking.



I have put hundreds of hours of work into this over the past half year to make it stable, reliable, and completely offline. If you want a lightweight memory engine that does not choke your home server, check it out on GitHub. Dropping a star on the repo helps a ton to get this noticed: [https://github.com/roandejager/Hillock](https://github.com/roandejager/Hillock)

document fetched 2026-10-29 00:31 · canonical https://reddit.com/r/selfhosted/comments/1xhjre9/after_6_months_of_daily_building_i_just_released