Embedded Memgraph: In‑Process Engine + Library API

[REASY]

Summary

Add an embedded mode for Memgraph that runs in‑process as a library, with a stable API for creating a graph instance, executing Cypher, and managing lifecycle without a server daemon.

Motivation / Use Cases

• CLI/TUI apps that want local, zero‑ops graph querying.
• Desktop apps with offline graph analytics.
• Unit tests and CI that need fast, isolated graphs.
• Edge/embedded environments where running a server is overkill.

Problems Today

• Memgraph only runs as a standalone server.
• Requires network setup, auth, process orchestration.
• Hard to embed into apps or use for fast local dev.

[gitbuda]

Yep, super reasonable request, + I would like to add that "agentic runtimes" would benefit massively out of this because there is a lot of in-process logic that would benefit (would be easier to implement the logic) if a solution like this existed (a subset of all "AI workloads").

If I may, what's missing under the existing solutions? like:

• https://github.com/kuzudb/kuzu (probably the best embedded engine, but the project "failed')
  • kuzu forks like:
    • https://github.com/predictable-labs/ryugraph
    • https://github.com/LadybugDB/ladybug
• https://github.com/lance-format/lance-graph (relatively recent/new project)

In other words, why would Memgraph go there, and what would it offer that is unique (there are solutions for all the above problems)?

In general, Memgraph's current architecture is not designed for the embedded use case because the team has always focused on the server-side use cases. In theory, we could offer a solution, but it's going to be a lot of work 🛠️

Btw, I've started doing some experimentation under memgraph/disk-vibes#1, it's far from being done or generally useful, any feedback 🙏

[REASY]

Hi, @gitbuda

First of all, Memgraph just works (at least for my workloads) and it is much less painful than neo4j. One of the most useful features for me that it supports most of the types I need: Maps, Lists, and primitives and that I can convert domain object to Cypher just by getting json and transforming it to QueryParam. This allowed me to quickly try out memgraph and model K8s objects without spending too much time.

I'm building a tool k8s-ariadne-rs, it is to query Kubernetes with natural language by compiling English to Cypher, it is powered by Memgraph, Rust, and LLMs. But to make that tool easier to use I decided to create a CLI version and CLI are supposed to be very easy to install and run. So I decided to try out graphqlite as embedded GraphDB, it sounded great, but in practice it isn't so good. Different engines, different behaviour. To make it somewhat working I had to do a lot of hacking like query rewrite.

It would have been much easier if I would rely on the same engine the would (hopefully) use the same parser and execution engine making me think less about GraphDB and more about what I'm building with it.

At this moment I decided to vibe a query engine on top of in-memory Petagraph (still a lot of work, parser, AST, query engine).

As why not kuzu forks, there are so many K8s objects that I'm just too lazy to think about their schema

[REASY]

@gitbuda thank you for showing lance-graph, I'll check it out. But one of the concerns I have after trying graphqlite, how well they implement at least openCypher and do they pass TCK.

[gitbuda]

Nice points and project! 👏 Also using https://github.com/memgraph/rsmgclient ❤️

[dylanbstorey]

@REASY ,

I see you’ve opened up a number of requests on graphqlite. I appreciate that you tried it out and are having some issues. The issues you opened i don’t think are unreasonable - if you can submit some actual code blocks to go with where you’re seeing problems I can get them moved into our test harness and get updates rolling pretty quickly.

, it sounded great, but in practice it isn't so good

I’ll be honest this one is a bit confusing to me as i’m using it regularly myself and at least agentic coders are quick to pick it up without problem so again “i’d like to here more about where you’re experiencing issues - is it the raw cypher or binding layers”.

Also PRs are welcome additions.

[REASY]

Hi, @dylanbstorey

Replying you here, colliery-io/graphqlite#13 (comment). At this point I decided to write openCypher parser + simple in-memory GraphDB over Petagraph (it is already used), https://github.com/REASY/k8s-ariadne-rs/tree/main/ariadne-cypher