Most apps that touch AI end up running three or four databases. A regular one for the records. A vector database for search by meaning. A graph database for the links between things. Maybe a key-value store for fast lookups. Then somebody writes the code that copies changes between them, and one day a copy goes missing. A deleted document still shows up in search. A link points at a record that’s gone.
HyperCrux keeps one copy. It’s a native hybrid of four databases, SQL, key-value, graph and vector, crossing at one record in one file. A record’s fields, its links and its vector change together in one transaction, so they can’t drift apart. You import it as a package in Go, or run the small hypercrux command and talk to it in JSON from any language. Backup is copying a file.
The Pitch in Four Numbers
- 4 in 1. SQL, key-value, graph and vector over the same records.
- 0 mismatches. A writer was killed mid-write 200 times, and records, links and vectors never fell out of step with the last committed transaction.
- 0.41 seconds to search 100,000 vectors of 384 values exactly, comparing every one, on a two-core cloud machine with every write flushed to disk. Among 10,000 vectors it takes about 42 milliseconds.
- Apache 2.0 for all of the code. Free to use, change and ship inside your own product.
How It Works
Every record has four ways in. Look it up by key. Query it with SQL. Walk its links to other records. Or find it by how similar its vector is to a question. All four hit the same row.
The 0.x line is built on SQLite, which does the hard parts: crash safety and locking. Records are rows in ordinary tables. The rules that tie keys, rows, links and vectors together are SQLite triggers stored inside the file, so even a plain SQL write from some other tool keeps everything consistent. Any language that can open SQLite can read the file directly, and the repository has a Python example that uses only the standard library.
Vector search is exact. It compares the question against every vector that passes the filter. That’s slower than an approximate index at large sizes, but it means filters and transactions just work, with no special cases and no index that can go stale.
The Beta Engine
Version 0.2 is the current stable line. The next one is already public. Since 7 October 2026 a new engine, the Beta, has been written in pure Go, split into small modular tasks and built by AI coding agents working in parallel. One week after the go-ahead it searched vectors 11 times faster than 0.1. It was tested with a power cut at every write and 261 planted bugs, and 0.3.0-beta.1 shipped on 15 October as a pre-release for Linux. Moving from 0.x takes an export of the data and a new import path for Go. ARM Linux is part of the Beta, and macOS comes after.
Who It’s For
Small teams that want AI features without running a cluster of services. The project site walks through five cases:
- A law firm. Twenty lawyers, thousands of contracts, and a clause library of similar clauses, precedents and client links that stays in the office.
- A PhD student. 1,400 papers in Zotero, and a literature review in one file with citation links, search over abstracts and SQL.
- A bike shop. About 3,000 products, with similar-product suggestions, parts that fit and stock levels side by side.
- A support bot. About 400 help articles, searched by meaning, and it never quotes one that’s been retired.
- An AI assistant’s memory. People, facts and embeddings stored in one transaction, so when something is forgotten it’s gone everywhere.
Limits, Stated Plainly
The project is upfront about where it stops. Search time grows with the number of vectors, and past about a hundred thousand vectors per table a database with an approximate index will answer faster. Several processes on one machine can share the file, but it has to sit on a local disk, because SQLite locking doesn’t work over network file systems. And HyperCrux stores and compares vectors without making them. You bring embeddings from the model of your choice.
The Four-Question Screening
Every project featured here gets the same four questions before anything else, the way you’d screen a startup idea.
1. Where is the money in it?
Not in the database itself. HyperCrux is free and open source, and today there’s no paid tier, no commercial licence and no support plan. That’s deliberate for now: a small embedded database earns trust by being free to try. The money, if it comes, sits around it. Support and long-term maintenance for teams that build products on it. A hosted or managed version for people who’d rather not run anything. Paid add-ons for the cases the free core leaves out, an approximate vector index above a hundred thousand vectors being the obvious first one. And the AI memory use case points at a larger market, since every assistant and agent product needs somewhere to keep people, facts and embeddings consistently.
2. What real problem does it solve?
Data that drifts apart across systems. When one set of facts lives in a regular database, a vector database and a graph database, every change has to be copied three times, and copies fail. The result is search results pointing at deleted documents, links to records that no longer exist, and a sync job someone has to babysit. For small teams the bigger cost is simply running three or four services where one would do.
3. The technology in plain words
Think of one index card per thing: a contract, a paper, a product. You can find the card by its number, by filtering on what’s written on it, by following the strings tied to other cards, or by asking which cards are most like this one. HyperCrux is the box that holds the cards, and it makes sure that when a card changes, its strings and its “likeness” change at the same moment. The box is one file on disk.
4. Moat and differentiation
The honest answer is that the moat is thin and has to be built. The code is open, so it can be copied. What sets HyperCrux apart is the combination: all four models given equal weight in one embedded file, one transaction across all of them, and crash testing that proves they stay in step. The big vector databases are separate servers, and the big general databases add vectors as an extension. Small, embedded and four-way consistent is a narrow spot that few others aim at. What would make it durable is a reputation for correctness built through the published test logs, a pure Go engine that’s easy to ship inside other programs, and the use-case library that shows people what to build with it.
Why It Matters
Vector search got bolted onto everything in the last two years. Mostly that meant a second system next to the real database and a sync job between them. HyperCrux bets that for a lot of apps the right answer is smaller: one file that does all four jobs and can’t fall out of step with itself. SQLite did that for ordinary tables a generation ago. If the Beta engine holds up, HyperCrux is the same move for AI-era data.
Follow the Project
Start with the quick start, get a build from the download page, or read the code in the HyperCrux repository on GitHub. The site also has a plain explainer on what a multi-model database is and the full write-up of how the Beta was tested. HyperCrux is listed with the other open-source projects on SoftDevStudio.com. Other featured projects here: AltSql and BareProxy.