Skip to content

OPENCLAW.md

Latest commit

 

History

History
210 lines (153 loc) · 5.19 KB

OPENCLAW.md

File metadata and controls

210 lines (153 loc) · 5.19 KB

RLM for OpenClaw

This branch adds OpenClaw compatibility to the Recursive Language Model (RLM) skill.

What is RLM?

RLM gives AI agents two capabilities:

  1. Analyze codebases beyond context limits - Recursively analyze large codebases using sub-agent delegation
  2. Persistent memory across sessions - Store and recall knowledge that survives restarts

Installation

Prerequisites

  • OpenClaw installed (brew install openclaw)
  • UV package manager (script will install if missing)

Install the Skill

# Clone the repo and run the install script (auto-detects OpenClaw path)
git clone https://github.com/marknutter/recursive-ai.git
cd recursive-ai
git checkout openclaw-compatibility
./install-openclaw.sh

If the install script cannot find your OpenClaw installation, set the path explicitly:

OPENCLAW_SKILLS_DIR=/path/to/openclaw/skills ./install-openclaw.sh

The install script will:

  1. Copy the skill file to OpenClaw's skills directory
  2. Install Python dependencies via UV
  3. Make the rlm CLI globally available

Usage

Mode 1: Code Analysis (Beyond Context Limits)

Analyze large codebases:

/rlm "find all security vulnerabilities" ./src

/rlm "summarize the architecture" ~/projects/myapp

/rlm "find uses of deprecated APIs" ./legacy-code

How it works:

  1. You (orchestrator) scan metadata and chunk the codebase
  2. Spawn sub-agents to analyze each chunk
  3. Sub-agents extract and read actual content (keeps it out of your context)
  4. You synthesize results without ever loading raw content

Benefits:

  • Analyze millions of lines of code
  • Stay under context limits
  • Parallel processing via sub-agents
  • Smart chunking strategies (by file, function, semantic units)

Mode 2: Persistent Memory

RLM includes MCP tools that are ALWAYS available:

Store Knowledge

// Tool call example (OpenClaw will expose these as functions)
rlm_remember({
  content: "Mark prefers Railway for cloud hosting. Uses it for Moxmo production.",
  tags: ["preferences", "infrastructure", "moxmo"],
  metadata: {
    source: "conversation-2026-02-19",
    importance: "medium"
  }
})

Recall Knowledge

rlm_recall({
  query: "hosting preferences",
  limit: 5
})
// Returns: Top 5 relevant memories with similarity scores

Browse Memories

rlm_memory_list({
  limit: 10,
  tags: ["moxmo"]
})
// Returns: 10 most recent entries tagged "moxmo"

Extract a Specific Memory

rlm_memory_extract({
  entry_id: "abc123"
})
// Returns: Full content of that memory entry

Delete a Memory

rlm_forget({
  entry_id: "abc123"
})

When to use memory:

Remember:

  • User preferences and patterns
  • Project decisions and rationale
  • Important facts that should persist
  • Context that spans multiple sessions

Recall:

  • At session start (what do I know about this user?)
  • Before recommendations (what are their preferences?)
  • When context is missing (did we discuss this before?)

Integration with Existing Memory

RLM memory complements OpenClaw's existing memory system:

  • MEMORY.md → Narrative context, curated insights (use memory_search tool)
  • Daily logs → Chronological session records
  • RLM memory → Structured, searchable, tagged knowledge

Use both together:

  1. RLM memory for quick facts, preferences, structured data
  2. MEMORY.md for narrative context, project history, relationships
  3. Daily logs for chronological session records

Architecture

OpenClaw-specific changes:

  1. Skill location: $(npm root -g)/openclaw/skills/rlm/ (auto-detected)
  2. CLI access: Global rlm command (via uv tool install)
  3. Sub-agent spawning: Uses sessions_spawn instead of Claude Code's Task tool
  4. MCP tools: Exposed via OpenClaw's tool system (not MCP servers)
  5. Metadata: OpenClaw-compatible frontmatter with requires and install sections

Differences from Claude Code Version

Feature Claude Code OpenClaw
Skill location ~/.claude/skills/rlm/ $(npm root -g)/openclaw/skills/rlm/ (auto-detected)
CLI prefix cd <path> && uv run rlm Global rlm command
Sub-agents Task tool sessions_spawn tool
MCP tools MCP server (stdio://rlm) Native OpenClaw tools
Installation ./install.sh ./install-openclaw.sh

Development

Test the skill:

# In an OpenClaw session
/rlm "test query" ./test-directory

Update the skill:

cd "$(npm root -g)/openclaw/skills/rlm/rlm-src"
git pull
uv sync

Uninstall:

rm -rf "$(npm root -g)/openclaw/skills/rlm"
uv tool uninstall rlm

Contributing

This OpenClaw branch should stay in sync with the main Claude Code version. When updating:

  1. Merge changes from main branch
  2. Test that OpenClaw-specific adaptations still work
  3. Update this README if behavior changes

Support

License

Same as main RLM project (see LICENSE file).