opensourceprojects.dev

A broadsheet for software that doesn't ask for your email

BM25 + kNN search, agent memory, and MCP in a single Rust binary
GitHub RepoViews13

Project Description

View on GitHub

What If Your Search Engine, Agent Memory, and MCP Server Were All One Binary?

You've probably got a search stack, a vector database, and some ad-hoc memory layer for your agents, each with its own config, its own upgrade path, and its own way of breaking. XERJ asks a different question: what if all of that lived in a single Rust binary with no JVM in sight? That's the pitch, and the README backs it up with a fairly specific set of capabilities.

What It Does

XERJ is a search and retrieval engine built in Rust and shipped as a single binary. It handles BM25 keyword search, kNN vector search, and hybrid queries that combine the two. On top of that it adds agent memory (exposed through a /__memory endpoint) and native MCP support, so agents can talk to it directly.

The architecture leans on zero-config autoindexing—you point it at your content and it indexes without a lot of ceremony. The default embedder is lexical and runs offline, which means you get vector search working out of the box without downloading anything. A neural embedder is available as an opt-in if you want it, but it pulls down roughly 90 MB. There's also a claimed ES-YAML conformance score of 1448/1458 with zero failures, which suggests the query layer is aiming at Elasticsearch-style compatibility rather than inventing its own dialect from scratch.

Why It's Cool

  • One binary, no JVM. If you've ever babysat an Elasticsearch cluster, the phrase "single binary, no JVM" probably landed with some weight. Fewer moving parts means fewer things to tune, restart, and explain to whoever's on call.

  • Sensible defaults on the embedder. Making the default embedder lexical and offline is a genuinely practical call. You don't need to download a model or reach for an API key just to get vector search running. The neural path is there when you want it, but it's opt-in rather than mandatory.

  • Agent memory is a first-class feature, not a plugin. The /__memory endpoint suggests the project treats agent state as part of the engine rather than something you bolt on with a separate service. That's a meaningful simplification if you're building agents and tired of stitching together a vector store plus a key-value store plus glue code.

  • Native MCP support. MCP has become the de facto way to connect models to tools, and having it built into the search engine means your agent can query your index without an adapter layer in between. The README also points to an llms.txt file, which is a nice touch for making the docs agent-readable.

  • Hybrid search without the assembly. BM25 plus kNN plus hybrid in one system is the combination most retrieval setups eventually converge on anyway. Getting it in one place saves you from reconciling scores across two different engines.

How to Try It

The README points to an install guide rather than listing commands inline, so start there. The general shape of getting going:

  1. Head to the repository and check the releases page for the binary for your platform.
  2. Follow the install docs to get it running.
  3. Point it at your content—autoindexing means there's no config file to write before your first query.
  4. Try a query against the query docs to see BM25, kNN, and hybrid modes in action.
  5. If you're wiring up an agent, look at the agentic memory recipe and the vectors docs for the embedder options.

There's also a Discord and GitHub Discussions if you hit a wall.

Final Thoughts

XERJ is aimed at people who want retrieval and agent memory without running a cluster—solo developers, small teams, anyone building agents who'd rather have one process than five. The single-binary approach and offline-by-default embedder are the kind of decisions that make a tool pleasant to actually deploy. It's still in release-candidate territory (the README references rc.101), so treat it as something to evaluate rather than bet production on without testing. But the scope is ambitious in a useful way, and if the pieces hold together, it could replace a fair amount of glue code in your stack.


Follow @githubprojects for more developer tools and open source projects.

Back to Projects
Project ID: xerjLast updated: November 17, 2026 at 06:53 AM