RLM - Recursive Language Model Context Tool

Prerequisites

IMPORTANT: RLM must be installed before use. Verify installation first:

bash

which rlm || dotnet tool install -g rlm

Requirements:

•.NET 10+ runtime

bash

# Install from NuGet
dotnet tool install -g rlm

# Update to latest version
dotnet tool update -g rlm

# Verify installation
rlm --version

Overview

RLM CLI implements the Data Ingestion Building Blocks pattern for processing documents that exceed your context window. It streams content using IAsyncEnumerable and maintains session state for multi-turn processing.

Use this skill when:

•Input exceeds your context window
•You need to find specific information in large documents (needle-in-haystack)
•You need to summarize or aggregate data from massive corpora
•You need to compare sections across large documents

Key limits:

•Max recursion depth: 5 levels (prevents infinite decomposition)

Quick Start

bash

# Ensure RLM is installed, then load document
dotnet tool install -g rlm 2>/dev/null; rlm load document.md
rlm chunk --strategy uniform --size 50000
# Process current chunk, then:
rlm store result_0 "extracted info"
rlm next                              # Get next chunk
rlm store result_1 "more info"
rlm aggregate                         # Combine all results

Documentation

Topic	File	Description
Agent Guide	agent-guide.md	Parallel processing with sub-agents
Strategies	strategies.md	All chunking strategies with decision tree
Examples	examples.md	Real-world workflow scenarios
Reference	reference.md	Complete command reference and JSON formats
Troubleshooting	troubleshooting.md	Common errors and solutions

Supported Formats

Format	Extension(s)	Features
Markdown	`.md`, `.markdown`	YAML frontmatter, code blocks, headers
PDF	`.pdf`	Text extraction, page count, title, author
HTML	`.html`, `.htm`	Converts to Markdown, preserves structure
JSON	`.json`	Pretty-prints, element count
Word	`.docx`	Heading preservation, paragraph extraction, document properties
Plain text	`.txt`, etc.	Basic text loading

Core Workflow

1. Load Document

bash

rlm load document.md                    # Single file
rlm load ./docs/                        # Directory (merged)
rlm load ./docs/ --pattern "**/*.md"    # Recursive glob
rlm load ./docs/ --merge false          # Keep separate
cat huge-file.txt | rlm load -          # From stdin

2. Check Document Info

bash

rlm info                # Size, tokens, metadata
rlm info --progress     # Processing progress bar

3. Choose Decomposition Strategy

Task	Strategy	Command
Find specific info	`filter`	`rlm filter "pattern"`
Summarize document	`uniform`	`rlm chunk --strategy uniform --size 50000`
Analyze structure	`semantic`	`rlm chunk --strategy semantic`
Token-precise	`token`	`rlm chunk --strategy token --max-tokens 512`
Complex documents	`recursive`	`rlm chunk --strategy recursive --size 50000`
Unknown task	`auto`	`rlm chunk --strategy auto --query "your question"`

See strategies.md for detailed options and selection guide.

4. Process Chunks

bash

rlm store chunk_0 "Finding from first chunk"
rlm next                                    # Get next chunk
rlm store chunk_1 "Finding from second chunk"
# Continue until "No more chunks"

5. Navigate Efficiently

bash

rlm skip 10       # Skip forward 10 chunks
rlm skip -5       # Skip backward 5 chunks
rlm jump 50       # Jump to chunk 50 (1-based)
rlm jump 50%      # Jump to 50% position

6. Aggregate Results

bash

rlm aggregate                       # Combine all stored results
rlm aggregate --separator "\n---\n" # Custom separator

7. Clear Session

bash

rlm clear         # Clear default session
rlm clear --all   # Clear all sessions

Parallel Processing

For documents with 10+ chunks, use parallel processing with sub-agents:

bash

# 1. Parent initializes with named session
rlm load massive.pdf --session parent
rlm chunk --strategy uniform --size 30000 --session parent

# 2. Parent extracts chunks and spawns workers
#    IMPORTANT: Chunking (step 1) must complete before using `next`
rlm next --raw --session parent > chunk_0.txt
# SPAWN: rlm-worker with "Process chunk_0.txt, session=child_0"

rlm next --raw --session parent > chunk_1.txt
# SPAWN: rlm-worker with "Process chunk_1.txt, session=child_1"
# ... continue for all chunks

# 3. After workers complete, import and aggregate
rlm import "rlm-session-child_*.json" --session parent
rlm aggregate --session parent

Alternative: Use slice for exporting content without chunking:

bash

# Export content by character position (no chunking required)
rlm slice 0:30000 --session parent --raw > chunk_0.txt
rlm slice 30000:60000 --session parent --raw > chunk_1.txt

Key Rules:

•Parent uses --session parent
•Each worker uses unique --session child_N
•Workers store results with key result

Recursive Delegation: Workers can spawn their own child workers for very large chunks. See agent-guide.md for the complete recursive delegation protocol.

Commands Quick Reference

Command	Description
`rlm load <file\|dir\|->`	Load document(s) into session
`rlm info [--progress]`	Show document metadata or progress
`rlm slice <range>`	View section (e.g., `0:1000`, `-500:`)
`rlm chunk [--strategy]`	Apply chunking strategy
`rlm filter <pattern>`	Filter by regex
`rlm next [--raw\|--json]`	Get next chunk
`rlm skip <count>`	Skip forward/backward
`rlm jump <index\|%>`	Jump to chunk index or percentage
`rlm store <key> <value>`	Store partial result
`rlm import <glob>`	Import child session results
`rlm results`	List stored results
`rlm aggregate`	Combine all results
`rlm clear [--all]`	Reset session(s)

For complete command options and JSON output formats, see reference.md.

Best Practices

•Start with info - Check document size before choosing strategy
•Filter first - For search tasks, use filter to reduce content
•Store incrementally - Save results after each chunk, not in batches
•Navigate efficiently - Use skip and jump instead of repeated next
•Merge small chunks - Use --min-size --merge-small for semantic chunking
•Clear between tasks - Run rlm clear when starting fresh

Permissions

This skill restricts tool access to Bash(rlm:*) only - Claude can only execute rlm commands when this skill is active.

For common errors and solutions, see troubleshooting.md.