Moradin Recall

moradin:recall

Search Moradin's memory and answer a question with citations.

When invoked

/moradin:recall <question>

or /moradin:recall <question> --topic <tag> for topic-filtered search.

What you do

1. Tokenize the question. Identify key terms.

2. Run BM25 search:

   python scripts/search.py "<question>" [--topic <tag>] [--applies-to <tag>] --limit 10
   

3. Read the top results (typically 5-10 files). For each, read both frontmatter and body.

4. Synthesize an answer that:

   • Directly addresses the question
   • Cites EACH source file by path (e.g. memory/patterns/closed_loop_eval.md)
   • Distinguishes between principles (universal rules), patterns (designs), references (external knowledge), and lessons (incident-derived)
   • Honest about gaps: if memory doesn't have the answer, say so

5. Optional: file the synthesis back. If the synthesis itself is valuable, ask the operator: "This synthesis is valuable — file it back as a new pattern or lesson?"

Citation format

Always cite specific files. Don't say "your memory has principles about closed loops" — say "see memory/principles/close_every_loop.md."

Topic filter

If the operator's question is clearly within one topic (e.g. "what do I know about evals?"), use --topic eval to focus the search.

If unsure: search without filter first; let BM25 surface relevant docs.

Don't

• Don't synthesize without reading the actual files. Search results give you hits — read them before answering.
• Don't make up content. If the memory doesn't say it, say "memory doesn't cover this."
• Don't filter results silently. If you skip a hit, explain why.

Output

• A synthesized answer with file-path citations
• Optionally: a new file in memory/patterns/ or memory/lessons/ if the synthesis is filed back