RLM (Recursive Language Model) Reasoning Pattern

Based on Zhang et al., MIT CSAIL (arXiv:2512.24601)

Core Principle: Context Outside, Not Inside

Never load entire large files into context. Instead, keep data "outside" and access selectively.

4-Step Reasoning Pattern

1. PROBE (Explore)

Search for relevant areas using keywords before reading:

code

Grep for keywords to find relevant lines
Glob to find related files by pattern

Example:

code

Grep(pattern="authentication", path="src/")
Glob(pattern="**/auth*.py")

2. EXTRACT (Load Selectively)

Only read the specific sections you need:

code

Read with offset/limit for specific line ranges
Load only chunks that contain matches

Example:

code

Read(file_path="src/auth/login.py", offset=50, limit=100)

3. CONFIRM (Verify)

Before answering, verify information against the source:

code

Re-check that found information matches the question
Validate references actually exist
Cross-reference multiple sources if needed

4. ANSWER (Respond)

Provide verified response based on confirmed evidence.

When to Apply This Pattern

Situation	Action
File > 1000 lines	PROBE first with Grep, never Read entire
Complex analysis	Delegate to Task tool sub-agent
Code generation	Verify imports/references exist with Grep
Multiple files	Use Glob to find, then selective Read

Anti-Patterns (Avoid)

code

# BAD: Loading entire file
Read("large_file.py")  # Don't do this!

# GOOD: Search first, then selective read
Grep("function_name", "src/")
Read("src/file.py", offset=150, limit=50)

Sub-Agent Delegation

For complex multi-step analysis, use Task tool:

code

Task(
    subagent_type="Explore",
    prompt="Find all authentication-related code and summarize the flow"
)

Verification Checklist

Before providing an answer:

• Did I search before reading?
• Did I read only necessary sections?
• Did I verify references exist?
• Is my answer based on confirmed evidence?

Additional Resources

•For Python implementation details, see rlm_agent.py
•For detailed API reference, see REFERENCE.md

Quick Commands

Run the RLM agent directly:

bash

python ~/.claude/skills/rlm-reasoning/scripts/rlm_query.py "your question" /path/to/file