Hillock

237 stars
AGPL-3.0
Active this week
Last commit today

Last verified Oct 29, 2026

Run a local neuro-symbolic memory engine for AI agents using HDC math, Hebbian plasticity, and SQLite, fitting under 1.2 GB VRAM.

Visit website GitHub237 Deploy on Hostinger

Read the full breakdownWho it is for, how it works, trade-offs, the quick start and the FAQ

who it's for

Who Hillock is for#

AI developers building local agent memory

Hillock provides persistent, factually grounded memory for agents built on Ollama models. Use it as the memory backend for a local assistant that reads documents via /ingest, builds a knowledge graph from them, and answers follow-up questions through the OpenAI-compatible API or directly via the IntegratedHillock Python library.

Skip if:

Skip if the agent requires multi-hop logical reasoning across many entities or multi-language document support. Both are v2.0 roadmap items not yet stable in v1.0.0.

Privacy-focused knowledge workers building a personal assistant

Ingest personal notes, text files, and PDFs via the /ingest command. Hillock builds a local SQLite knowledge graph from the content, connects entities through Hebbian associations, and lets you query it through Open-WebUI or Obsidian without any data leaving the machine.

Skip if:

Skip if you need a polished no-code interface out of the box. Setup requires Python, Ollama, and basic CLI comfort. A connected Open-WebUI instance provides a browser UI after initial setup.

Researchers prototyping neuro-symbolic AI memory

The AGPL-3.0 source exposes the full HDC/VSA implementation, Hebbian weight updates, and graph triple storage in Python. The verify_hillock.py test suite validates the math on CPU without a GPU or large model weights. Fork and modify the similarity gate threshold, encoding scheme, or graph traversal logic freely under AGPL terms.

Skip if:

Skip if you need production-ready multi-hop reasoning for evaluation. The v2.0 resonator network work for multi-hop queries is in the research roadmap, not yet in the stable release.

Teams evaluating local RAG alternatives for edge deployments

Hillock's sub-1.2 GB VRAM footprint and Ollama compatibility make it testable on a standard developer laptop. Point the API at Open-WebUI to get a chat interface over the local knowledge graph. The CPU-only fallback means testing does not require a GPU-equipped machine.

Skip if:

Skip if you need production-ready multi-user deployments or enterprise SLAs. v1.0.0 is a single-user local tool. Enterprise deployments require the commercial license and direct contact per the README.

the problem

The problem it solves#

Most local RAG setups for AI agents are overbuilt for the hardware they run on. A typical pipeline loads a vector database, an embedding model, and a generative model to retrieve and synthesize facts, with three processes competing for GPU memory. On consumer hardware with a limited VRAM budget, the memory layer competes directly with the reasoning model.

The deeper challenge is accuracy. Vector similarity search finds roughly related text chunks and passes them to the LLM whether or not those chunks contain the answer. The LLM then synthesizes a response from partial evidence and produces confident wrong answers on anything outside the ingested corpus. Teams running local agents for personal notes, private documents, or domain-specific knowledge cannot afford hallucinated facts in retrieved memory.

how Hillock solves it

How it solves it#

HDC Similarity Gate

Before calling the LLM, Hillock computes a Hyperdimensional Computing (HDC) vector similarity score between the query and its knowledge base. If the score falls below the configured threshold, the engine returns that it does not know rather than passing the query to the model. This hard gate is the primary defense against hallucinated answers on out-of-scope queries and is evaluated entirely on local data.

SQLite Knowledge Graph

Facts are stored as semantic triples in a local SQLite database rather than raw text chunks. Hillock extracts clean entity relationships during ingestion and stores them as graph nodes with typed connections. An alias command merges duplicate entity references (for example, two names for the same person) without a separate deduplication pass.

Hebbian Synaptic Associations

As facts are added and queries are answered, Hillock updates synaptic weights between related graph nodes following Hebbian plasticity rules. Nodes that activate together strengthen their association over time, which improves retrieval specificity on follow-up questions about related entities. The /inspect command shows current weights for any stored node.

OpenAI-Compatible Local API

The bundled FastAPI server runs at http://127.0.0.1:8000 and exposes an OpenAI-compatible endpoint at /v1. Connect Open-WebUI, AnythingLLM, or Obsidian by pointing them at the local URL as a custom OpenAI base URL. No connector code required. The terminal console and a Python library (IntegratedHillock) provide two additional integration modes using the same local database.

Under 1.2 GB VRAM

The engine fits under 1.2 GB VRAM on a consumer GPU and falls back to CPU when no NVIDIA card is present. Small, specialized bi-encoder models handle fact extraction during ingestion instead of a full generative model, keeping memory overhead low. A 30-sentence text file ingests in roughly 5 seconds per the README.

Three Integration Modes

Hillock runs as a background API server for external apps, an interactive terminal console for direct chat and commands (/ingest, /inspect, /alias, /model, /mode), or as an importable Python library via IntegratedHillock from the engine module. Each mode reads from and writes to the same local SQLite database file.

strengths · trade-offs

Strengths and trade-offs#

Strengths

Offline with No Outbound API CallsEvery retrieval and inference step runs on local hardware with no external network requests. Memory content never leaves the machine, which matters for agents handling personal notes, private documents, or sensitive business data. This contrasts with cloud-based agent memory APIs that require sending query context to a remote server on every interaction.
Hallucination-Resistant by ArchitectureThe HDC similarity gate is a structural mechanism, not a prompt-engineering workaround. When the knowledge base lacks relevant facts, the engine refuses to answer rather than confabulating. The behavior is inspectable: the /inspect command shows stored facts and synaptic weights for any entity, so failures are traceable to specific gaps in the knowledge graph.
Edge Hardware TargetThe sub-1.2 GB VRAM footprint and CPU-only fallback make Hillock usable on laptops and mini PCs that typical RAG stacks cannot run. Ollama handles the generative tier; Hillock provides the memory tier without adding significant GPU memory pressure on top of the model already loaded.
Inspectable Memory StateThe /inspect command shows all stored facts, synaptic weights, and graph structure for any named entity. Unlike opaque vector databases where retrieved chunks are invisible until queried, every node in Hillock's graph is readable and auditable from the terminal at any time.

Trade-offs

-Multi-Step Setup Before First RunGetting Hillock running requires installing Python 3.10+, installing Ollama separately, downloading an Ollama model, installing hillock via pip, and downloading the en_core_web_sm SpaCy model. There is no Docker image that bundles these dependencies. Users who have not set up a Python virtual environment or Ollama before will need to work through each prerequisite in sequence.
-English-Only Fact Extraction in v1.0.0The entity parser uses the en_core_web_sm SpaCy model, trained on English-language text. Ingesting documents in other languages will produce degraded fact extraction quality. Multi-language support is not documented in v1.0.0, and no alternative language model is specified in the README.
-AGPL Copyleft Requires Commercial License for Closed AppsUnder AGPL-3.0, running a modified version of Hillock as a network service requires releasing the modified source. Embedding it in a closed-source application or proprietary SaaS product without open-sourcing that code requires the commercial license at $49 per month through Polar. Self-hosters running Hillock for internal team use without distributing or serving a modified version can use the AGPL tier at no cost.

versus alternatives

Hillock vs alternatives#

Hillock vs Mem0

Mem0 is a managed cloud service that adds persistent memory to AI agents through an API. Hillock serves the same use case but runs entirely on local hardware with no external API calls.

Feature	Hillock	Mem0
License	AGPL-3.0	Proprietary
Deployment	Local Python process	Cloud API
Offline operation	Yes	No
Data privacy	All data stays local	Stored on Mem0 servers
Hallucination handling	Hard HDC similarity threshold	Retrieval confidence scoring
Cost model	Free (AGPL); $49/mo commercial	Free tier; paid plans by usage

Hillock is the better fit when offline operation, data privacy, or VRAM constraints are the deciding factors. Because retrieval and inference run locally, there are no per-request API costs and no memory content sent to a third-party server. Mem0's managed API is the simpler path for developers who want zero infrastructure and can accept cloud-side storage of agent memory context.

Hillock vs Zep

Zep provides long-term memory for AI assistants with a self-hosted Community Edition (Apache 2.0) and a paid cloud platform. Like Hillock, Zep can run on-premises, but its memory architecture uses traditional vector embeddings with a graph layer, while Hillock uses HDC math and Hebbian plasticity as the primary retrieval mechanism.

Feature	Hillock	Zep
License	AGPL-3.0	Apache 2.0 (CE) / Proprietary (cloud)
Memory model	Neuro-symbolic HDC + SQLite graph	Vector embeddings + graph
VRAM requirement	Under 1.2 GB	Depends on embedding model chosen
Hallucination handling	Hard HDC gate (refuses unknown queries)	Retrieval with confidence scoring
Multi-user support	Single-user local	Multi-user (cloud and CE)

Hillock's hard similarity gate gives a binary response on out-of-scope queries, which suits agents where a wrong answer is worse than no answer. Zep's Community Edition is more mature, has broader SDK support across languages, and handles multi-user deployments. For single-user local agents on constrained hardware, Hillock's smaller footprint is the practical advantage. For production teams serving multiple users or needing stable SDK integration, Zep's Community Edition or managed cloud is the more established path.

install · quick start

Quick start#

bash

Install via pip with Python 3.10+ and a running Ollama instance as prerequisites.
```bash
pip install hillock
python -m spacy download en_core_web_sm
```

tech stack · detected from GitHub

What it's built on#

Languages: Python
Frameworks: FastAPI

frequently asked

FAQ#

Does Hillock require a GPU to run?

No. Hillock falls back to CPU when no NVIDIA GPU is available. The under 1.2 GB VRAM figure is the GPU footprint when a card is present; the engine runs without one. The Ollama instance Hillock delegates inference to may have its own GPU preferences, but many Ollama models run acceptably on CPU.

What license does Hillock use?

Hillock is dual-licensed under AGPL-3.0 and a commercial license. The AGPL-3.0 tier is free for self-hosted personal and internal team use. Running a modified version as a network service requires releasing modified source under AGPL terms. Embedding Hillock in a closed-source application or SaaS product without open-sourcing that code requires the commercial license at $49 per month through Polar.

How does Hillock prevent hallucinated answers?

Before sending a query to the LLM, Hillock computes an HDC (Hyperdimensional Computing) vector similarity score between the query and the knowledge base. If the score falls below the threshold, the engine returns that it does not have the answer instead of passing the query to the model. This is a deterministic check against the local database, not a confidence estimate generated by the LLM itself.

Which apps can connect to Hillock?

Hillock runs a local FastAPI server at http://127.0.0.1:8000 with an OpenAI-compatible endpoint at /v1. Any application that accepts a custom OpenAI base URL can connect to it. The README names Open-WebUI, AnythingLLM, and Obsidian as tested integrations.

How do I ingest documents into Hillock?

From the terminal console (python main.py), use the /ingest command with a local .txt or .pdf file path. Hillock uses small bi-encoder models to extract facts, stores them as semantic triples in the SQLite knowledge graph, and builds Hebbian associations between related entities. Per the README, a 30-sentence text file takes roughly 5 seconds to ingest.

also worth a look

Similar open-source tools#

WeKnora

Alternative toGlean

Self-hosted knowledge base with RAG search, AI agent, and wiki

35,598 GitHub stars
Licence MIT

Qdrant

Alternative toPinecone

Self-hosted vector database for AI similarity search and RAG

35,258 GitHub stars
Licence Apache-2.0

Weaviate

Alternative toPinecone

AI-native vector database for semantic search and AI apps

16,916 GitHub stars
Licence BSD-3-Clause

Hermes Agent

Featured

Alternative toGrok Bot

Self-hosted AI agent: persistent memory, multi-channel chat, model choice.

257,359 GitHub stars
Licence MIT

Gpt Researcher

Alternative toPerplexity AI

Deep research agent for cited reports from web and local documents.

30,339 GitHub stars
Licence Apache-2.0

Happy

InDeveloper Tools

Open source AI coding orchestrator for Claude, Codex, Grok, Gemini.

24,386 GitHub stars
Licence MIT

by category

Open source tools in these categories#

423

Self-host Hillock on a Hostinger VPS

Keep it running 24/7, even when your laptop is off. A Hostinger VPS gives it a server of its own.

01
Choose a plan
Starts from $6.49/mo
02
Apply
Extra 10% off yearly plans
03
Install Hillock
On a server that stays on
04
Start building
Focus on the app, not the server

Deploy on Hostinger

Inbox Zero

Alternative toSuperhuman

AI email assistant that drafts replies and clears your inbox

12,826 GitHub stars
Licence AGPL-3.0
#Email

Omnara

Alternative toClaude Managed Agents

Open-source agent deployment API. Self-host or use Omnara Cloud.

3,296 GitHub stars
Licence Apache-2.0
#AI Agents

rakazo

Alternative toGrok Bot

AI teammates you own: your keys, your model, your machine.

5,914 GitHub stars
Licence Apache-2.0
#AI Agents

Advertise your product

The projects we add and every new guide. Free, unsubscribe any time.

Find open source alternatives to paid software

Hillock vs Mem0

Mem0 is a managed cloud service that adds persistent memory to AI agents through an API. Hillock serves the same use case but runs entirely on local hardware with no external API calls.

Feature	Hillock	Mem0
License	AGPL-3.0	Proprietary
Deployment	Local Python process	Cloud API
Offline operation	Yes	No
Data privacy	All data stays local	Stored on Mem0 servers
Hallucination handling	Hard HDC similarity threshold	Retrieval confidence scoring
Cost model	Free (AGPL); $49/mo commercial	Free tier; paid plans by usage

Hillock vs Zep

Feature	Hillock	Zep
License	AGPL-3.0	Apache 2.0 (CE) / Proprietary (cloud)
Memory model	Neuro-symbolic HDC + SQLite graph	Vector embeddings + graph
VRAM requirement	Under 1.2 GB	Depends on embedding model chosen
Hallucination handling	Hard HDC gate (refuses unknown queries)	Retrieval with confidence scoring
Multi-user support	Single-user local	Multi-user (cloud and CE)

Who Hillock is for#

AI developers building local agent memory

Skip if:

Privacy-focused knowledge workers building a personal assistant

Skip if:

Researchers prototyping neuro-symbolic AI memory

Skip if:

Teams evaluating local RAG alternatives for edge deployments

Skip if:

The problem it solves#

How it solves it#

HDC Similarity Gate

SQLite Knowledge Graph

Hebbian Synaptic Associations

OpenAI-Compatible Local API

Under 1.2 GB VRAM

Three Integration Modes

Strengths and trade-offs#

Strengths

Trade-offs

Hillock vs alternatives#

Hillock vs Mem0

Hillock vs Zep

Quick start#

What it's built on#

FAQ#

Similar open-source tools#

Open source tools in these categories#

AI & Machine Learning

Databases & Storage

LLMOps & AI Tooling

Find open source alternatives to paid software

Who Hillock is for#

AI developers building local agent memory

Skip if:

Privacy-focused knowledge workers building a personal assistant

Skip if:

Researchers prototyping neuro-symbolic AI memory

Skip if:

Teams evaluating local RAG alternatives for edge deployments

Skip if:

The problem it solves#

How it solves it#

HDC Similarity Gate

SQLite Knowledge Graph

Hebbian Synaptic Associations

OpenAI-Compatible Local API

Under 1.2 GB VRAM

Three Integration Modes

Strengths and trade-offs#

Strengths

Trade-offs

Hillock vs alternatives#

Hillock vs Mem0

Hillock vs Zep

Quick start#

What it's built on#

FAQ#

Similar open-source tools#

Open source tools in these categories#

AI & Machine Learning

Databases & Storage

LLMOps & AI Tooling

Find open source alternatives to paid software