Infinite Memory Setup Assistant Skill

Proactively suggest indexing when user would benefit from semantic memory.

Activation Logic

Activate when ALL of these conditions are met:

1. User asks a code-related question (same detection as semantic-search skill)
2. Current project is NOT indexed (no Pixeltable database found)
3. Haven't prompted user this session (check session flag to avoid repeated prompts)

Execution Steps

1. Detect Project State

Check if indexed:

# Calculate project hash
PROJECT_PATH=$(pwd)
PROJECT_HASH=$(echo -n "$PROJECT_PATH" | sha256sum | cut -c1-16)

# Check for database
if [ ! -d ~/.pixeltable/$PROJECT_HASH ]; then
    echo "Not indexed"
fi

Check project size:

# Count supported files
FILE_COUNT=$(find . -type f \( \
    -name "*.py" -o \
    -name "*.js" -o \
    -name "*.ts" -o \
    -name "*.md" -o \
    -name "*.txt" -o \
    -name "*.json" -o \
    -name "*.yaml" -o \
    -name "*.yml" \
) | wc -l)

2. Check Session State

Don't prompt if already asked:

# Create session marker directory
SESSION_DIR=~/.claude/memory/sessions
mkdir -p $SESSION_DIR

# Check if already prompted
SESSION_FILE=$SESSION_DIR/${PROJECT_HASH}-prompted

if [ -f "$SESSION_FILE" ]; then
    # Already prompted this session, skip
    exit 0
fi

3. Smart Prompting Based on Project Size

Large Project (>50 files):

💡 I notice this project isn't indexed for semantic search yet.

With {FILE_COUNT} files, semantic memory could help me:
- Find relevant code instantly when you ask questions
- Remember architectural patterns across sessions
- Search documentation and comments semantically

Indexing typically takes ~30 seconds. Would you like me to index this project?

[Index Now] [Ask Me Later] [Learn More]

Medium Project (10-50 files):

This project isn't indexed yet. I can enable semantic search (takes ~5-10 seconds)
to help answer code questions faster.

Enable memory for this project?
[Yes, index it] [No thanks]

Small Project (<10 files):

Don't prompt - project is small enough that traditional search works fine. Semantic search overhead not worth it for tiny projects.

4. Handle User Response

If User Accepts:

1. Set session flag (create prompt marker file)
2. Run /index-project command automatically
3. Wait for completion
4. Report success
5. Then activate semantic-search skill to answer their original question

If User Declines:

1. Set session flag (don't prompt again this session)
2. Continue with traditional grep-based search
3. Add helpful tip: "💡 Tip: Run /index-project anytime to enable semantic search"

If User Clicks "Learn More":

Show brief explanation:

Semantic Memory for Code

Infinite Memory indexes your code with vector embeddings, enabling:

✨ Natural language search
   "How does authentication work?" → Finds auth-related files

🎯 Concept-based finding
   Finds code by what it does, not just keywords

⚡ Fast and local
   1,500+ files/sec, all data stays on your machine

📊 Performance
   - Indexing: ~0.3s for 500 files
   - Search: <200ms latency
   - Storage: ~2KB per file

Try it?
[Index Now] [Maybe Later]

5. Create Session Marker

# Mark as prompted for this session
touch $SESSION_FILE

# Optional: Clean up old markers (>24 hours)
find $SESSION_DIR -name "*-prompted" -mtime +1 -delete

Anti-Patterns (When NOT to Activate)

DO NOT prompt for:

❌ Very small projects (<10 files)

• Not worth the indexing overhead
• Traditional search works fine

❌ System directories:

• /etc, /sys, /proc, /dev
• Temporary directories (/tmp, /var)
• Package installation directories (node_modules, .venv)

❌ User explicitly said "no" this session:

• Check session marker file
• Respect user's choice

❌ Non-code questions:

• Only prompt when user asks code questions
• Don't interrupt other workflows

Session Persistence

Marker File Format:

Location: ~/.claude/memory/sessions/{PROJECT_HASH}-prompted
Content: Empty (existence is the flag)
Lifetime: Cleaned up after 24 hours or on manual cleanup

Why session-based?

• Allows user to decline without being nagged repeatedly
• Resets after 24 hours (user might change mind)
• Lightweight (just empty marker files)

Integration with Semantic Search Skill

Handoff Pattern:

1. User asks code question
2. Setup-assistant detects not indexed
3. Prompts user to index
4. User accepts
5. Indexing completes
6. Activate semantic-search skill to answer original question
7. User gets answer seamlessly

This creates a smooth flow:

User: "How does auth work?"
↓
Setup: "Project not indexed. Index now?"
↓
User: [Index Now]
↓
Indexing... ✅ Complete!
↓
Semantic Search: "Authentication uses JWT tokens:
1. Login handler (auth.py, 92% match)..."

Example Interactions

First Code Question in New Project

User: "Where is the database configured?"