recursive-language-model（递归语言模型）

RLM (Recursive Language Model) Skill

This skill enables processing of arbitrarily long documents by treating them as an external environment and recursively calling sub-LLMs over chunks of the content.

Overview

The RLM pattern follows the architecture from the paper:

• Root Agent: Main agent orchestrating the task
• Sub-LLM (llm_query): Subordinate agent for chunk-level analysis
• External Environment: Persistent Python REPL with document state

Requirements

• Python 3.8+
• An agentic harness with code execution and subordinate agent capabilities

File Structure

rlm_skill/
├── SKILL.md           # This file
├── rlm-subcall.md     # Subordinate agent prompt profile
└── scripts/
    └── rlm_repl.py    # Persistent Python REPL

Quick Start

1. Initialize the skill by loading it into your agent context
2. Load a context file using the REPL init command
3. Run the RLM workflow - chunk, delegate, synthesize

Detailed Usage

Step 1: Initialize the REPL with your context

Execute via your agent's code execution tool:

python3 /path/to/rlm_skill/scripts/rlm_repl.py init /path/to/your/context.txt

Step 2: Verify status

python3 /path/to/rlm_skill/scripts/rlm_repl.py status

Step 3: Scout the context

python3 /path/to/rlm_skill/scripts/rlm_repl.py exec -c 'print(peek(0, 3000))'

Step 4: Create chunks

python3 /path/to/rlm_skill/scripts/rlm_repl.py exec -c 'print("\\n".join(write_chunks("./rlm_chunks", size=200000, overlap=0)))'

Step 5: Delegate to sub-LLM

Use your agentic harness's subordinate agent capability with the rlm-subcall profile:

Subordinate Agent Query:
Profile: rlm-subcall
Message: Query: <user_question>

Chunk file: ./rlm_chunks/chunk_0000.txt

Extract relevant information from this chunk.

Step 6: Synthesize results

Combine sub-LLM outputs and provide the final answer in the main conversation.

REPL Helper Functions

The rlm_repl.py provides these injected functions:

• peek(start=0, end=1000) - View portion of content
• grep(pattern, max_matches=20, window=120, flags=0) - Search with context
• chunk_indices(size=200000, overlap=0) - Get chunk boundaries
• write_chunks(out_dir, size=200000, overlap=0, prefix='chunk') - Materialize chunks as files
• add_buffer(text) - Store intermediate results

REPL Commands

• init <context_path> - Load context file
• status [--show-vars] - Show current state
• exec [-c "code"] - Execute Python code (reads from stdin if no -c)
• export-buffers <out_path> - Write buffers to file
• reset - Delete state file

Best Practices

1. Don't paste large chunks into main chat - Use REPL and subordinates
2. Use grep() to locate relevant sections - Then peek() for details
3. Keep chunk outputs structured - Prefer JSON for sub-LLM responses
4. Synthesize in main conversation - After collecting evidence
5. Clean up state when done - Use reset to clear old contexts

Example Workflow

User: "Analyze this 500KB log file and find all error messages from service X"

Agent:
1. Init: python rlm_repl.py init /path/to/logs.txt
2. Scout: exec -c 'print(peek(0, 500))'
3. Search: exec -c 'print(grep("service X.*error", max_matches=10))'
4. Chunk: exec -c 'write_chunks("./chunks", size=100000)'
5. For each chunk, call subordinate with rlm-subcall profile
6. Synthesize findings and answer user

Limitations

• State is stored locally via pickle (not distributed)
• Context size limited by available memory
• Sub-LLM calls are synchronous in this implementation

References

• MIT paper: "Recursive Language Models"
• Implementation: brainqub3/claude_code_RLM