# Hillock

> A lightweight, 100% local neuro-symbolic memory engine combining hyperdimensional computing, Hebbian plasticity, and SQLite knowledge graphs for offline AI agents.

Hillock is an open-source, local memory engine built by Roan de Jager for edge hardware and privacy-first AI workflows. It replaces vector databases and LLM-based extraction pipelines with a three-tier neuro-symbolic architecture that runs entirely offline, requiring less than 1.2 GB of VRAM. The project is licensed under AGPL-3.0 and is actively developed on GitHub, currently at version 0.6.0.

## What It Is

Hillock is a local agent memory system that answers queries from a deterministic knowledge graph rather than approximate vector similarity. It ingests text and PDF documents using the TALON extraction pipeline — a three-stage process combining coreference resolution (Fastcoref), a bi-encoder predicate router (MiniLM), and a zero-shot relation extractor (GLiREL Large) — without ever calling a generative LLM during ingestion. A local LLM via Ollama is invoked only once a query has already passed a hard similarity gate, ensuring the model only generates responses grounded in verified facts.

## Three-Tier Memory Architecture

Hillock's memory system is composed of three complementary layers that work together:

- **SQLite Knowledge Graph**: Stores ground-truth facts as Subject-Predicate-Object (SPO) triples in relational tables, eliminating vector drift and approximation errors for factual recall.
- **Hebbian Plasticity Engine**: Tracks co-occurring concepts across conversation turns using gradient-free synaptic learning (η=0.15, decay γ=0.01) to mimic associative memory, surfaced live via the `/inspect` command.
- **Hyperdimensional Reservoir (VSA/HDC)**: A 10,000-dimensional bipolar hypervector space that compresses conversation context with fading-memory decay, resolves pronouns, binds multi-hop relational paths via positional permutations, and hard-blocks unanswerable queries in under a millisecond.

## HYDRA Gating and Hallucination Prevention

The HYDRA (Bipolar Late-Interaction MaxSim Gating) system is the core anti-hallucination mechanism. It replaces legacy single-vector query bundling with token-level MaxSim scoring natively in 10,000-D bipolar space. A Sub-Dimensional Projection Cascade evaluates MaxSim across a 2,000-D slice first to early-reject approximately 95% of candidate facts in ~0.5ms. The gate is a hard cutoff at `HDC_THRESHOLD = 0.55` — queries that do not clear this threshold receive a hardcoded refusal ("I do not have verified information about that.") rather than a hallucinated answer, consuming zero GPU generation cycles.

## Edge Hardware Design and Performance

Hillock was explicitly designed for mid-range and low-power hardware. The README reports the following benchmark comparisons between standard local RAG and Hillock's TALON+HDC pipeline:

- **Ingestion latency** for a 30-sentence document: ~5 seconds at 6.3 sentences/sec (GPU) vs. 15–30 minutes for an 8B LLM pipeline.
- **VRAM footprint**: under 1.2 GB (tested on a GTX 1070) vs. 5.8–16 GB+ for LLM-based extraction.
- **CPU-only execution**: fully supported; the complete ingestion and benchmark suite runs on a standard laptop CPU with no GPU requirement.
- **Sub-second retrieval**: the 2,000-D projection cascade evaluates 32 queries in approximately 1.16 seconds (~0.036s per query) on CPU.

## Update: v0.6.0 — The HyperGraph & Late-Interaction Expansion

Version 0.6.0, published on August 29, 2026, is described in the README as the largest architectural update to Hillock to date. Key changes include:

- **HYDRA late-interaction gating** replacing legacy single-vector bundling, eliminating query-length decay.
- **HYPERGRAPH-HDC multi-hop path reasoning** using Positional Permutation (cyclic coordinate shifts) to break Hadamard binding commutativity, enabling 2-hop and 3-hop relational reasoning during ingestion without LLM calls.
- **Hybrid Edge Storage Schema** constraining SQLite strictly to 1-hop SPO triples while storing multi-hop paths as bit-packed BLOBs in a dedicated `hdc_reservoirs` table.
- **21-point verification suite** (`verify_hillock.py`) upgraded with permutation orthogonality and sequential path validation checks, safe to wire into CI without a GPU runner.

Benchmark results for v0.6.0 show extraction recall of 59.1%, hard-negative block rate of 60.0%, and retrieval accuracy of 54.5% on the project's current fixed benchmark of one 32-sentence academic text with 22 answerable and 10 hard-negative queries. The README explicitly notes these numbers are directional rather than statistically robust, with a larger benchmark planned before any v1.0 claims.

## Setup and CLI Workflow

Hillock ships with one-click launcher scripts (`run.bat` for Windows, `run.sh` for Linux/macOS) that create a virtual environment, install dependencies, and check for the spaCy `en_core_web_sm` model automatically. Manual setup requires Python 3.10+, PyTorch with optional CUDA support, and a locally running Ollama instance with at least one instruct-tuned model. The interactive CLI supports commands for document ingestion (`/ingest`), personality mode switching (`/mode strict|balanced|conversational`), live model switching (`/model`), entity inspection (`/inspect`), and debug verbosity control (`/debug`).

## Features
- 100% local, offline execution with zero cloud dependencies
- Three-tier memory: SQLite knowledge graph, Hebbian plasticity engine, and 10,000-D HDC reservoir
- TALON extraction pipeline (Fastcoref + MiniLM + GLiREL Large) without LLM ingestion
- HYDRA bipolar late-interaction MaxSim gating to prevent hallucinations
- Multi-hop relational path reasoning via positional permutation (2-hop and 3-hop)
- Sub-dimensional projection cascade for sub-millisecond query rejection
- Hard similarity gate with hardcoded refusal for unanswerable queries
- Streaming response generation via local Ollama LLM
- Interactive CLI with /ingest, /inspect, /mode, /model, /debug commands
- 21-point GPU-free verification suite (verify_hillock.py)
- CPU-only execution supported
- Under 1.2 GB VRAM footprint
- One-click launcher scripts for Windows and Linux/macOS
- AGPL-3.0 open-source license with CLA for contributors

## Integrations
Ollama (local LLM inference), SQLite, spaCy (en_core_web_sm), Fastcoref (coreference resolution), MiniLM (bi-encoder predicate routing), GLiREL Large (zero-shot relation extraction), PyTorch (CUDA and CPU), GloVe embeddings (50-D, ~50K vocabulary)

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, CLI

## Pricing
Open Source

## Version
v0.6.0

## Links
- Website: https://github.com/roandejager/Hillock
- Repository: https://github.com/roandejager/Hillock
- EveryDev.ai: https://www.everydev.ai/tools/hillock