Releases·Tool Release
Hillock v0.6.0 adds ColBERT-style MaxSim late interaction to 10,000-D HDC edge memory engine
A developer shipped Hillock v0.6.0, a local AGPL-3.0 neuro-symbolic memory engine for edge hardware, and posted a technical breakdown of its retrieval and gating architecture. The release replaces bundled query hypervectors with ColBERT-style late-interaction (MaxSim) scoring over 10,000-dimensional bipolar hyperdimensional vectors. It targets CPU-only or <1.2 GB VRAM edge deployments, combining an SQLite knowledge graph, Vector Symbolic Architectures, and Hebbian associative memory.
11 engagement1 sourceSun, Aug 30, 2026, 02:26 PM
AI Overview
- Hillock v0.6.0 ships a retrieval and gating architecture adapting ColBERT's late-interaction MaxSim to 10,000-dimensional bipolar hyperdimensional computing
- Earlier versions superposed (bundled) query tokens into a single hypervector, which the new late-interaction approach replaces
- Hillock is a local, AGPL-3.0 neuro-symbolic memory engine built for edge hardware running CPU-only or under 1.2 GB VRAM
- The stack combines an SQLite knowledge graph, 10,000-D Vector Symbolic Architectures (VSA/Hyperdimensional Computing), and Hebbian associative memory instead of standard dense embeddings
Related
Products·launched
Hillock — neuro-symbolic local document memory engine replacing vector DBs and LLM extraction with SQLite, Hebbian links, and HDC gating
Hillock is an experimental AGPL-3.0 FOSS neuro-symbolic memory engine for local document retrieval. It replaces dense vector stores and LLM-based extraction passes with SQLite storage, Hebbian links, and hyperdimensional computing gating. It targets edge and consumer constraints where RAG ingestion is too slow and dense cosine similarity cannot reliably reject out-of-domain queries.
Hillock is an experimental AGPL-3.0 FOSS neuro-symbolic memory engine for local document retrieval. It replaces dense vector stores and LLM-based extraction passes with SQLite storage, Hebbian links, and hyperdimensional computing gating. It targets edge and consumer constraints where RAG ingestion is too slow and dense cosine similarity cannot reliably reject out-of-domain queries.
20 engagement2 sourcesredditMon, Sep 28, 2026, 07:23 AM
Products·building
Hillock: offline neuro-symbolic memory engine for local AI, an alternative to vector databases
Hillock is a lightweight, fully offline memory engine for edge hardware intended to replace vector databases in local AI setups. It uses a deterministic, neuro-symbolic approach (with SQLite) so the LLM can retrieve known facts without hallucinating when information is missing. It targets hobbyists running local AI on modest hardware like a GTX 1070 or basic CPU laptop.
Hillock is a lightweight, fully offline memory engine for edge hardware intended to replace vector databases in local AI setups. It uses a deterministic, neuro-symbolic approach (with SQLite) so the LLM can retrieve known facts without hallucinating when information is missing. It targets hobbyists running local AI on modest hardware like a GTX 1070 or basic CPU laptop.
6 engagement1 sourceredditSun, Aug 30, 2026, 02:25 PM
Products·building
Hillock: local neuro-symbolic memory engine for Ollama on <1.2 GB VRAM
Hillock is a local memory engine for Ollama that replaces vector-database RAG with a neuro-symbolic store, running on low-end edge hardware like a GTX 1070 or CPU-only laptops. A non-LLM component (TALON) extracts knowledge triples locally, so the LLM is only used for generation and can abstain instead of guessing. It targets users running 8B+ models on budget GPUs who need memory without burning VRAM.
Hillock is a local memory engine for Ollama that replaces vector-database RAG with a neuro-symbolic store, running on low-end edge hardware like a GTX 1070 or CPU-only laptops. A non-LLM component (TALON) extracts knowledge triples locally, so the LLM is only used for generation and can abstain instead of guessing. It targets users running 8B+ models on budget GPUs who need memory without burning VRAM.
9 engagement1 sourceredditSun, Aug 30, 2026, 02:24 PM
Releases·Tool Release
Hillock v0.4 local neuro-symbolic memory engine released
Developer roandejager released v0.4 of Hillock, a local neuro-symbolic memory engine for agents that runs on mid-range hardware. It uses SQLite SPO knowledge graphs and hyperdimensional vectors instead of LLM parsing during ingestion, reducing VRAM usage to under 1.2GB. The update adds schema type constraints and fixes inverted facts.
Developer roandejager released v0.4 of Hillock, a local neuro-symbolic memory engine for agents that runs on mid-range hardware. It uses SQLite SPO knowledge graphs and hyperdimensional vectors instead of LLM parsing during ingestion, reducing VRAM usage to under 1.2GB. The update adds schema type constraints and fixes inverted facts.
22014 engagement4 sourcesredditThu, Aug 13, 2026, 05:42 PM
Products·building
Hillock: local memory engine using SQLite triples and VSA vectors
Hillock is a local memory and RAG tool that replaces vector databases with a neuro-symbolic architecture: a SQLite knowledge graph for factual triples and Hebbian learning for connections, plus 10,000-D VSA vectors. It runs on modest hardware like a GTX 1070, aiming to reduce hallucination and VRAM usage for local LLM applications.
Hillock is a local memory and RAG tool that replaces vector databases with a neuro-symbolic architecture: a SQLite knowledge graph for factual triples and Hebbian learning for connections, plus 10,000-D VSA vectors. It runs on modest hardware like a GTX 1070, aiming to reduce hallucination and VRAM usage for local LLM applications.
4 engagement2 sourcesredditSun, Aug 23, 2026, 08:03 PM