Rlm

Recursive Language Model (RLM)

What is RLM

Standard LLMs degrade on long inputs — accuracy drops as the context window fills up ("context rot"). RLM avoids this by never loading full content into the prompt. Instead, the model receives only a manifest (a list of files and line counts) and navigates the content on demand using structured tags.

The model drives its own exploration: it peeks at specific lines, delegates sub-questions to recursive sub-calls, and emits an answer when it has enough information. This mirrors how an expert developer reads an unfamiliar codebase — skimming structure, drilling into relevant sections, ignoring noise.

When to use

• Analysing a large codebase for a cross-cutting concern ("find all places that touch auth")
• Summarising or querying a long document (research paper, log file, data export)
• Any question where the relevant content is spread across many files or sections
• When a direct prompt exceeds the model's context limit

Not needed for: small files that fit comfortably in a single prompt.

Prerequisites

• Python 3.9+
• A working --llm-cmd (see below)

Quick start

Query a codebase:

python skills/rlm/rlm.py \
  --query "Find all places that handle authentication" \
  --input ./src \
  --llm-cmd "claude -p"

Query a document:

python skills/rlm/rlm.py \
  --query "Summarise the key findings" \
  --input report.txt \
  --llm-cmd "python skills/rlm/prompt.py"

Query from stdin:

cat notes.txt | python skills/rlm/rlm.py \
  --query "What action items are mentioned?" \
  --input - \
  --llm-cmd "ollama run llama3"

Filter files in a large repo:

python skills/rlm/rlm.py \
  --query "Find all SQL queries" \
  --input ./backend \
  --include "*.py" \
  --exclude "*/migrations/*" \
  --llm-cmd "claude -p"

The --llm-cmd contract

--llm-cmd is any shell command that:

• reads the prompt from stdin
• writes the response to stdout

Examples:

# Claude Code
--llm-cmd "claude -p"

# Reference OpenAI adapter (included in this skill)
export OPENAI_API_KEY=sk-...
--llm-cmd "python skills/rlm/prompt.py"

# Ollama (local model)
--llm-cmd "ollama run llama3"

# Custom wrapper script
--llm-cmd "./my_agent.sh"

rlm.py itself satisfies the contract — you can use it as a --llm-cmd inside another caller to chain recursive processing.

Tuning

When limits are reached, the best partial answer accumulated so far is returned.

Pattern matching notes

File filters use Python's fnmatch. Key behaviour:

• *.py matches at any depth (e.g. src/auth/login.py) — * absorbs /
• tests/* only covers one level deep — tests/sub/foo.py is NOT excluded
• ** glob syntax is NOT supported
• For depth-independent matching, omit slashes: test_*.py excludes test files anywhere