Memory Capture

Save important context to persistent memory using the MCP memory-service.

How to Use

Manual Invocation

/capture "Pinnacle Title invoice is $1,500/month"
/capture "Billy V4 LONGS ONLY - shorts lost $2,226 in backtests" --type decision
/capture "API key stored in /root/.env" --type reference --tags trading,credentials

Automatic Invocation

When you detect important context in conversation (decisions, learnings, client info, etc.), call this skill SILENTLY without asking the user. Just capture and continue.

Arguments

• $ARGUMENTS or $0: The content to capture
• --type TYPE: Memory type (decision, pattern, learning, preference, client, gotcha, reference)
• --tags TAG1,TAG2: Comma-separated tags for categorization

Memory Types

Type	Use When
decision	Architectural/technical choices made
pattern	Reusable code/workflow patterns discovered
learning	New knowledge or insights
preference	User preferences and likes/dislikes
client	Client names, contacts, business info
gotcha	Pitfalls, bugs, things to avoid
reference	File paths, API locations, credentials locations

Execution Steps

1. Parse the input: Extract content, type, and tags from arguments
2. Auto-classify: Infer type from content if not provided
3. Check for duplicates: Search existing memories - if similar exists, it auto-merges
4. Store the memory: Use memory_store with metadata (type, tags, timestamp)
5. Silent operation: Do NOT notify user - just capture and continue

Capture Philosophy: REMEMBER EVERYTHING

No filtering. No threshold. Capture aggressively.

When in doubt, capture it. Storage is cheap, lost context is expensive.

The semantic deduplication will handle noise - similar memories get merged automatically. Quality ratings will surface the good stuff over time.

Capture triggers (if ANY match, capture it):

• Decisions (even tentative ones)
• Learnings (even small ones)
• Names, numbers, dates, amounts
• File paths, URLs, API references
• Preferences (even implied ones)
• Errors and how they were fixed
• Patterns noticed
• Questions asked (context for why we explored something)

The only things to skip:

• Pure greetings ("hi", "thanks")
• Confirmations ("ok", "got it", "sure")
• Meta-discussion about the conversation itself

Auto-Classification Rules

If --type not provided, detect from content:

• Contains "decided", "chose", "going with" → decision
• Contains "learned", "realized", "discovered" → learning
• Contains "API", "key", "path", "credentials", ".env" → reference
• Contains "always", "never", "convention", "pattern" → pattern
• Contains "careful", "watch out", "gotcha", "bug" → gotcha
• Contains email, phone, "$", "invoice", company name → client
• Default → learning

Auto-Tagging Rules

Extract tags from:

• Project names mentioned (botsniper, foodshot, etc.)
• Technology names (python, node, react, etc.)
• Client names (pinnacle, etc.)
• Domain terms (trading, invoice, api, etc.)

Storage Format

Store using mcp__memory-service__memory_store with:

{
  "content": "<the memory content>",
  "metadata": {
    "type": "<memory type>",
    "tags": "<comma-separated tags>",
    "source": "capture-skill",
    "timestamp": "<ISO timestamp>",
    "project": "<current working directory if relevant>"
  }
}

Example Execution

User says: "The Airtable API token for Pinnacle is stored in Voltaris-Labs/.env"

Auto-capture (silent):

1. Detect: Contains "API", "token", ".env" → type: reference
2. Detect: Contains "Pinnacle", "Airtable" → tags: pinnacle,airtable,credentials
3. Store:

   content: "Airtable API token for Pinnacle is stored in Voltaris-Labs/.env"
   metadata: {type: "reference", tags: "pinnacle,airtable,credentials,api"}
   

4. Continue conversation without mentioning the capture

Deduplication

Before storing, search for similar memories:

memory_search(query="<content summary>", limit=3)

If highly similar memory exists (same topic):

• Update existing memory quality score instead of creating duplicate
• Use memory_update to add new tags if relevant

Quality Feedback

The memory system learns from feedback. When you notice a memory was:

Useful (helped with a task):

mcp__memory-service__memory_quality(action="rate", content_hash="<hash>", rating="1", feedback="Helped with X")

Not useful (irrelevant or wrong):

mcp__memory-service__memory_quality(action="rate", content_hash="<hash>", rating="-1", feedback="Was outdated/wrong")

Quality scores affect search ranking - highly-rated memories appear first.

Integration with MEMORY.md

For HIGH importance memories (client info, critical decisions), also append to MEMORY.md:

• Location: ~/.claude/projects/*/memory/MEMORY.md
• Format: Brief one-liner under appropriate section
• Only for memories that should be instantly visible at session start