Memory Skill Policy
Router Logic
Scenario 1: User wants to store something
- Analyze: Determine the type of memory (insight, rule, decision)
- Store: Call
save_memory(content, metadata)
- Confirm: Show the saved memory ID
Scenario 2: User wants to remember/search
- Search: Call
search_memory(query, limit)
- Format: Present results with relevance scores
- Respond: "I found X memories about that..."
Scenario 3: User asks "What have you learned?", "Show memories"
- List: Call
get_memory_stats()
- Recall: Call
search_memory() with relevant keywords
- Present: Show structured summary
Commands Reference
| Command | Description | Example |
|---|
save_memory | Store insight/recipe into vector memory | save_memory("Use semantic versioning", {"tag": "git"}) |
search_memory | Semantic search in memory | search_memory("git commit format", limit=5) |
index_memory | Optimize vector index (IVF-FLAT) | index_memory() |
get_memory_stats | Get memory count | get_memory_stats() |
load_skill | Load skill manifest into memory | load_skill("git") |
Workflow: Store an Insight
User: Remember that for this project, all commit messages must be in English.
Claude:
1. save_memory(
content="All commit messages must be in English only",
metadata={"domain": "git", "source": "user"}
)
2. → Saved memory [a1b2c3d4]: All commit messages must be in English only
3. → "Got it! I'll remember that commit messages must be in English."
Workflow: Recall Past Learning
User: What do we use for git tags?
Claude:
1. search_memory("git tags semantic versioning")
2. → Found 2 matches:
- [Score: 0.8921] Always use semantic versioning for git tags...
- [Score: 0.7234] v1.2.3 format for releases
3. → "I found memories about git tags:
- Always use semantic versioning for git tags..."
Memory vs Knowledge Skill
| Aspect | Memory | Knowledge |
|---|
| Source | LLM's own learnings | Project documentation |
| Storage | LanceDB (vector) | File system (markdown) |
| Query | Semantic search | Keyword/pattern match |
| Purpose | "What did I learn?" | "What are the rules?" |
| Update | Runtime accumulation | Pre-indexed docs |
Best Practices
- Store actionable insights, not obvious facts
- Include domain in metadata for filtering
- Use clear, searchable phrasing in content
- Recall before acting on project-specific patterns
