Today I'm releasing Minigraf 1.0 — a single-file embedded graph database written in Rust, with Datalog as its query language and bi-temporal semantics built in.
If you've ever wanted Datomic but embedded, or SQLite but for connected data with full history, this is that. One .graph file, no server, no clustering, no Java runtime. It runs natively, in WebAssembly, on Android, on iOS, and inside agent processes.
The shape of the problem
Most agent memory today is one of two things: a vector store, or a pile of JSON files. Both are wrong for the same reason — they don't model the relational structure of what an agent knows, and they don't preserve when it knew it.
A coding agent that decided last Tuesday to use Postgres over MongoDB needs to remember not just the decision, but the rationale, the alternatives considered, and what the codebase looked like at the time. A vector store can find a similar-sounding chunk; it can't reconstruct the decision graph. A JSON pile can store the data; it can't query "show me every decision motivated by performance concerns made before the v2 refactor."
That second query — graph traversal scoped by time — is what Minigraf is built for.
What 1.0 actually ships
The core query model is Datalog over EAV triples (entity-attribute-value), which is the same model Datomic and XTDB use. Minigraf's contribution is the packaging:
- Bi-temporal by design. Every fact carries both transaction time (when it was recorded) and valid time (when it was true in the world). Retract a fact and the history is preserved — you can replay the database state at any past moment.
-
Single file, embedded.
Minigraf::open("data.graph")and you're running. No daemon, no port, no config. The entire database is one file you can ship, version, or sync. - Recursive Datalog rules. Multi-hop graph traversals are native, not bolted on. Reachability, transitive closure, and path queries are one rule each.
-
Prepared statements. Parse and plan once, execute thousands of times with
$slotbind tokens. Roughly the same ergonomics assqlite3_prepare_v2. -
Window functions in
:find.sum,count,min,max,avg,rank,row_numberwith:over (partition-by … :order-by …). Aggregations without dropping out of Datalog. -
Mobile and WASM. Native Kotlin (
.aar) and Swift (.xcframework) bindings via UniFFI. Browser WASM viawasm32-unknown-unknown, server WASM viawasm32-wasip1.
use minigraf::{Minigraf, OpenOptions};
let db = OpenOptions::new().path("agent.graph").open()?;
db.execute(r#"
(transact [[:decision-42 :decision/chose "postgres"]
[:decision-42 :decision/considered "mongodb"]
[:decision-42 :motivated-by :perf-concern]])
"#)?;
// What did we believe at transaction 100?
db.execute(r#"
(query [:find ?choice
:as-of 100
:where [?d :decision/chose ?choice]
[?d :motivated-by :perf-concern]])
"#)?;
What it isn't
Minigraf is deliberately not distributed, not cluster-scale, and not client-server. It's optimised for sub-million-node graphs that live next to a single process — an agent, a mobile app, a browser tab. If you have a billion nodes and a fleet of services, you want XTDB or Neo4j. If you have an agent that needs to remember a month of decisions and replay any past state, you want Minigraf.
It's also not a time-series database. Time-series tools (InfluxDB, Prometheus) are optimised for measurements at a point in time. Minigraf is optimised for facts with full retraction history — different problem, different shape.
Why now
The bi-temporal-graph thesis isn't new. Datomic has had it since 2012; XTDB built on the idea; my own previous project, RecallGraph, was pursuing the same vision in 2019. What's new is the audience.
Agents are the use case bi-temporal databases were waiting for. An agent's memory has all the structural properties — entities with relationships, beliefs that get corrected, decisions with provenance, states that need to be reconstructed at past moments — that bi-temporal Datalog was designed for. The vector store wave gave us "find similar," but agents that reason need "what did I believe and why." Minigraf is built for that second question.
How to try it
[dependencies]
minigraf = "1.0"
Or:
cargo add minigraf
The repo is at github.com/project-minigraf/minigraf. The README has a quickstart, the wiki has a Datalog reference and a comparison page (vs XTDB, Cozo, Neo4j, SQLite), and there's a working agent memory skill at github.com/project-minigraf/temporal_reasoning that demonstrates the bi-temporal model end to end.
If you build something on Minigraf, I want to hear about it — open a Discussion on the repo or reach out directly. The next phase of the project is shaped by what people actually use it for, and 1.0 is when that conversation starts.
⭐ Star the repo if the thesis resonates — it helps surface the project to others working on similar problems.
