RLM-RS Plugin for Claude Code

A Claude Code plugin that integrates the rlm-rs Rust CLI for processing documents that exceed typical context window limits using the Recursive Language Model (RLM) pattern.

Overview

This plugin enables Claude Code to:

• Process documents up to 100x larger than the context window
• Use multiple chunking strategies (fixed, semantic, parallel)
• Hybrid semantic + BM25 search to find relevant chunks efficiently
• Pass-by-reference chunk retrieval (no file I/O overhead)
• Delegate chunk-level analysis to efficient sub-LLM calls
• Synthesize coherent answers from distributed chunk analyses

Prerequisites

rlm-rs CLI must be installed:

# Via Cargo (recommended)
cargo install --git https://github.com/zircote/rlm rlm-rs

# Via Homebrew
brew install zircote/tap/rlm-rs

Installation

From Marketplace

# Add the zircote marketplace (if not already added)
claude plugin marketplace add zircote https://github.com/zircote/marketplace

# Install the plugin
claude plugin install rlm-rs@zircote

From GitHub

claude plugin install zircote/rlm-rs-plugin

For Development

# Use --plugin-dir flag for local development
claude --plugin-dir /path/to/rlm-rs-plugin

Verify Installation

Run /help in Claude Code - you should see rlm-rs:rlm-init, rlm-rs:rlm-load, etc.

Components

Skills

• rlm: Main RLM workflow orchestration for large context processing
• rlm-chunking: Guidance on selecting chunking strategies

Commands

Command	Description
/rlm-init	Initialize the RLM database
/rlm-load	Load a file into a buffer with chunking
/rlm-status	Show current buffers and state
/rlm-query	Run a query against loaded content

Agents

• rlm-subcall: Haiku-based agent for efficient chunk analysis (uses pass-by-reference with chunk IDs)
• rlm-synthesizer: Sonnet-based agent for aggregating JSON findings into coherent responses

Usage

Basic Workflow

1. Initialize the database:

   /rlm-init
   

2. Load a large file:

   /rlm-load file=path/to/large-document.txt chunker=semantic
   

3. Query the content (uses hybrid semantic + BM25 search):

   /rlm-query query="What are the main themes discussed?"
   

The query command automatically:

• Searches for relevant chunks using hybrid search
• Generates embeddings on first search (cached for future queries)
• Analyzes only the most relevant chunks
• Synthesizes a coherent answer

Using the RLM Skill

For automated orchestration, trigger the RLM skill by asking:

• "Process this large file using RLM"
• "Analyze document.txt which exceeds context"
• "Use recursive language model to examine logs.txt"

Configuration

Create .claude/rlm-rs.local.md in your project for custom settings:

---
chunk_size: 6000
overlap: 0
default_chunker: semantic
---

Architecture

Based on the RLM paper (arXiv:2512.24601):

RLM Concept	Implementation
Root LLM	Main Claude Code conversation (Opus/Sonnet)
Sub-LLM	rlm-subcall agent (Haiku)
External Environment	rlm-rs CLI with SQLite storage

Contributing

See CONTRIBUTING.md for development setup and guidelines.

License

MIT