rlm

rlm-go

A Go implementation of Recursive Language Models (RLM) - an inference-time scaling strategy that enables LLMs to handle arbitrarily long contexts by treating prompts as external objects that can be programmatically examined and recursively processed.

Overview

RLM-go provides a Go REPL environment where LLM-generated code can:

• Access context stored as a variable
• Make recursive sub-LLM calls via Query() and QueryBatched()
• Use standard Go operations for text processing
• Signal completion with FINAL() or FINAL_VAR()

Key Design Decisions

Unlike the Python RLM which uses socket IPC, rlm-go uses direct function injection via Yaegi - a Go interpreter. This eliminates:

• Socket server overhead
• Serialization/deserialization
• Process boundaries

The result is ~100x less latency per sub-LLM call compared to socket IPC.

Requirements

• Go 1.23 or later (for building from source)
• An LLM API key:
  • ANTHROPIC_API_KEY for Claude models (default)
  • GEMINI_API_KEY for Gemini models
  • OPENAI_API_KEY for OpenAI models
• (Optional) Podman or Docker for isolated sandbox execution

Supported Models

Provider	Models	Env Variable
Anthropic	claude-sonnet-4-20250514, claude-opus-4-20250514, etc.	ANTHROPIC_API_KEY
Google	gemini-3-flash-preview, gemini-3-pro-preview	GEMINI_API_KEY
OpenAI	gpt-5, gpt-5-mini	OPENAI_API_KEY

The provider is auto-detected based on model name. Anthropic is the default.

Installation

Quick Install (Recommended)

# Download and install the latest release
curl -fsSL https://raw.githubusercontent.com/XiaoConstantine/rlm-go/main/install.sh | bash

This installs the rlm binary to ~/.local/bin/rlm.

Go Install

go install github.com/XiaoConstantine/rlm-go/cmd/rlm@latest

From Source

git clone https://github.com/XiaoConstantine/rlm-go.git
cd rlm-go
go build -o rlm ./cmd/rlm

As a Library

go get github.com/XiaoConstantine/rlm-go

Claude Code Integration

RLM includes a skill for Claude Code that provides documentation and usage guidance for large context processing.

Install the Skill

rlm install-claude-code

This creates a skill at ~/.claude/skills/rlm/SKILL.md that teaches Claude Code:

• When to use RLM (contexts >50KB, token efficiency needed)
• Command usage and options
• The Query() and FINAL() patterns
• Token efficiency benefits (40% savings on large contexts)

After installation, restart Claude Code to activate the skill.

CLI Usage

# Basic usage with Anthropic (default)
rlm -context file.txt -query "Summarize the key points"

# Use Gemini
rlm -model gemini-3-flash-preview -context file.txt -query "Analyze this data"

# Use OpenAI
rlm -model gpt-5-mini -context file.txt -query "Summarize this"

# Verbose output with iteration details
rlm -context logs.json -query "Find all errors" -verbose

# JSON output for programmatic use
rlm -context data.csv -query "Extract anomalies" -json

# Pipe context from stdin
cat largefile.txt | rlm -query "What patterns do you see?"

CLI Options

Flag	Description	Default
-context	Path to context file	-
-context-string	Context string directly	-
-query	Query to run against context	Required
-model	LLM model to use	claude-sonnet-4-20250514
-max-iterations	Maximum iterations	30
-verbose	Enable verbose output	false
-json	Output result as JSON	false
-log-dir	Directory for JSONL logs	-

Quick Start (Library)

package main

import (
    "context"
    "fmt"
    "log"
    "os"

    "github.com/XiaoConstantine/rlm-go/pkg/rlm"
)

func main() {
    // Create your LLM client (implements rlm.LLMClient and repl.LLMClient)
    client := NewAnthropicClient(os.Getenv("ANTHROPIC_API_KEY"), "claude-sonnet-4-20250514")

    // Create RLM instance
    r := rlm.New(client, client,
        rlm.WithMaxIterations(10),
        rlm.WithVerbose(true),
    )

    // Run completion with long context
    result, err := r.Complete(context.Background(), longDocument, "What are the key findings?")
    if err != nil {
        log.Fatal(err)
    }