oneiron

Embedded retrieval engine for memory-first applications.

One binary. One process. Zero network hops.

Why

Most retrieval stacks bolt together separate services for vectors, text, and graphs — network hops, consistency gaps, and operational complexity that doesn't belong on a phone. Oneiron runs in-process as a Rust library with C FFI bindings. Every query touches a single LMDB environment with ACID transactions. Embed it on iOS, Android, desktop, or a server.

Signals

Signal	Engine	What it finds
Vector	HNSW (flat NSW), SIMD-accelerated	Semantically similar content
Text	BM25 inverted index	Exact keywords and phrases
Graph	Personalized PageRank over typed edges	Relationally connected entities
Temporal	Bi-temporal range indexes	Events by when they happened or were recorded
Phonetic	Code-based posting lists	Fuzzy matches from voice/ASR misspellings

Any subset of signals can be combined via Reciprocal Rank Fusion with per-signal boosts.

Quick Start

use oneiron::{Vault, VaultConfig, EntityId};

let config = VaultConfig::device();
let vault = Vault::open("./my-vault", config)?;

let id = EntityId::now();
vault.put_entity(&id, b"msgpack blob")?;
vault.put_vector(&id, &embedding)?;

Building

cargo build --release
cargo test

Design

18 LMDB databases per vault. Atomic multi-database writes via BatchBuilder. MessagePack entity blobs. Context packing into LLM-ready formats.

Full details in the design docs:

• SCHEMA-DESIGN.md — database layout, key formats, encoding
• BUILD-PROMPT.md — architecture, algorithms, API surface
• DEPLOYMENT.md — multi-vault deployment, ML infrastructure

License

Apache 2.0