RLM (Recursive Language Model)

RLM is Gee-Code’s mode for handling complex tasks that benefit from persistent state, multi-step computation, and sandboxed code execution. When enabled, the AI gets a Python sandbox where it can execute code, store variables, and build up results iteratively.

How RLM Works

In standard mode, the AI uses tools (Read, Write, Edit, Bash) to interact with the filesystem. In RLM mode, the AI additionally gets a persistent Python REPL:

Standard Mode:
  AI -> Tool calls -> File system -> Results -> AI

RLM Mode:
  AI -> Tool calls + Python sandbox -> Persistent state -> AI
      |-- rlm_exec: Execute Python code
      |-- rlm_get_var: Read sandbox variables
      |-- rlm_chunk: Split large data for processing
      |-- llm_query: Send chunks to a sub-LLM for analysis
      |-- llm_batch: Process multiple chunks in parallel

The Sandbox

The sandbox provides an isolated Python execution environment:

• Persistent variables — variables survive across multiple exec calls
• Standard library access — json, re, math, collections, itertools, datetime
• File I/O through tools — file writes go through the tool system (not direct disk access)
• Sub-LLM calls — send content to a smaller, faster model for analysis
• Chunking — split large content at meaningful boundaries for parallel processing

When to Use RLM

RLM is most useful for:

• Data transformation — processing large datasets, CSV manipulation, JSON restructuring
• Complex analysis — tasks requiring multi-step reasoning with intermediate state
• Large file processing — chunking and analyzing files too big for a single context window
• Code generation — building up code programmatically (templates, scaffolding)
• Context gathering — /context gather uses RLM for intelligent file discovery

Enabling RLM

/rlm on           # Enable RLM mode
/rlm off          # Disable
/rlm              # Toggle

When enabled, all subsequent messages use the RLM execution path. The status line shows when RLM is active.

Per-Bead RLM

Individual beads can enable RLM without turning it on globally:

Bead(
    title="Process data files",
    agent="code",
    rlm_enabled=True
)

RLM Tools

Tool	Description
rlm_exec	Execute Python code in the sandbox
rlm_get_var	Read the value of a sandbox variable
rlm_chunk	Split content into chunks (by size, lines, pattern, or separator)
llm_query	Send a chunk + query to a sub-LLM for analysis
llm_batch	Process multiple chunks in parallel
rlm_complete	Signal early completion when the answer is found

Chunking Strategies

Large content can be split intelligently:

• strategy="size" — line-boundary aware chunks (default 8000 chars)
• strategy="lines" — fixed line count chunks
• strategy="pattern" — split at code boundaries (class/function definitions)
• strategy="separator" — split at a separator string (paragraphs)

RLM Display

RLM operations show animated progress with timing:

> Local Analysis (12.3s)
|-- Step 1: Thinking (2.1s)
|   |-- Read: app.tsx (0.3s)
|   |-- Glob: **/*.tsx (0.1s)
|-- Step 2: Processing (1.2s)
|   |-- rlm_exec: transform_data() ...
|-- Step 3: Pending

RLM vs Standard Mode

Aspect	Standard Mode	RLM Mode
State	Stateless between tool calls	Persistent Python sandbox
Execution	Tool calls only	Tool calls + Python REPL + sub-LLM
Best for	File editing, search, simple tasks	Complex computation, data processing
Token usage	Lower per interaction	Higher (multi-step sandbox)

Next Steps

• Planning & Execution — task decomposition with beads
• Tools Overview — the complete tool list
• Context Management — /context gather uses RLM