fatby/rlm-engine

RLM Engine — Explicit deep reasoning over history

IMPORTANT: Explicit activation only

This skill ONLY activates when the user EXPLICITLY requests RLM analysis. Do NOT use this skill based on question complexity — let the user decide.

Trigger phrases (user must say one of these)

• /rlm
• use RLM to...
• analyze with RLM...
• RLM search...
• deep analysis of...

Examples of activation

✓ "/rlm how many times did we discuss Docker?" 
✓ "Use RLM to find patterns in our conversations" 
✓ "Analyze with RLM what topics we covered this month" 
✓ "Deep analysis of all infrastructure decisions" 

✗ "What did we talk about yesterday?" → Use memory_search (no RLM trigger) 
✗ "Find patterns in our chats" → Use memory_search (no explicit RLM request) 
✗ "Summarize our project discussions" → Use memory_search (no explicit RLM request) 

Why explicit activation?

| Tool | Speed | Cost | When to use | |------|-------|------|-------------| | memory_search (native) | ~100ms | Free | Default for all memory questions | | rlm-engine (this skill) | 15-180s | $0.01-0.10 | Only when user explicitly asks |

The user knows when they need deep analysis and will request it explicitly. Do not try to guess — if they don't say "RLM", use memory_search.

What RLM can do that memory_search cannot

• Execute Python code to count, aggregate, calculate statistics
• Iterate over ALL sessions (not just top-K matches)
• Cross-reference data across many sessions programmatically
• Detect patterns that require reading complete transcripts

Invocation

1. Confirm activation: "Running RLM deep analysis... this can take 15-180 seconds depending on context and model iterations."
2. Execute:

cd {baseDir} && uv run python src/rlm_bridge.py \ 
  --profile-model cost \ 
  --query "EXACT USER QUESTION (without the /rlm prefix)" 

3. Handle result JSON:
   • status: "ok" → respond with response field
   • status: "rate_limited" → Kimi API limit, suggest waiting or using memory_search
   • status: "skipped" → not enough history
   • status: "error" → inform problem, suggest memory_search as fallback

Technical details

• Models (default profile cost): kimi-k2.5 (root) + kimi-k2.5 (sub) + kimi-k2-turbo (fallback)
• Context: loads up to 30 sessions, 200K characters max (configurable)
• API: Moonshot API (requires MOONSHOT_API_KEY environment variable)
• Cost: ~$0.01-0.05 per query (pay-as-you-go)