/recall

name: recall
description: Search past PRDs and decisions with full-text and semantic search
category: Memory

Recall Command

Search project history for past PRDs, decisions, and learnings.

Purpose

Enable quick access to historical context:

• Find similar past features ("How did we implement OAuth?")
• Recall technical decisions ("What database schema did we use?")
• Learn from past blockers ("What issues came up with Stripe integration?")
• Discover reusable patterns ("Show me all API implementations")

Usage

Basic Search (Full-Text)

/recall "OAuth implementation"
/recall "Stripe payment"
/recall "database schema"

Filters

# Search only completed PRDs
/recall "authentication" --status=complete

# Search by priority
/recall "feature" --priority=P0

# Search by date range
/recall "api" --since=30d
/recall "bug" --before=2025-10-01

# Search specific PRD
/recall "performance" --prd=PRD-003

Semantic Search (Optional)

# Find similar features (requires embeddings)
/recall --similar-to=PRD-003
/recall --similar-to="real-time chat"

Workflow

Step 1: Initialize Memory Index (First Run)

On first use, create SQLite FTS5 index:

→ Initializing search index...
→ Scanning product/prds/...
  • Found 25 PRDs
  • Indexing content...
  • Building FTS5 index...
✓ Index created (.claude/memory/index.db)

This will speed up future searches significantly.

Index updates automatically when PRDs change.

Step 2: Full-Text Search

Search using SQLite FTS5:

SELECT
  prd_id,
  name,
  status,
  priority,
  snippet(prds_fts, -1, '<mark>', '</mark>', '...', 30) AS snippet,
  rank
FROM prds_fts
WHERE prds_fts MATCH 'oauth OR authentication'
ORDER BY rank
LIMIT 10;

Step 3: Rank and Display Results

Show most relevant PRDs:

🔍 Search Results for "OAuth implementation"

Found 3 matches:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📅 PRD-003: User Authentication System
✅ Status: Complete | P0 | Completed 2 months ago

🔑 Key Decisions:
  • Used OAuth 2.0 with PKCE flow
  • Chose NextAuth.js over Passport.js
  • Token refresh every 30 minutes
  • Stored refresh tokens in HttpOnly cookies

⚠️ Blockers Encountered:
  • CORS issues with Google OAuth (fixed with proxy)
  • Token rotation edge cases (solved in v1.1)
  • Safari third-party cookie restrictions (used workaround)

💡 Learnings:
  • Always test token expiration scenarios
  • Document OAuth redirect URLs clearly
  • Consider token security vs UX tradeoffs

📎 Related PRDs:
  • PRD-007: SSO Integration (similar pattern)
  • PRD-012: API Authentication (depends on this)

📄 File: product/prds/04-complete/241015-user-auth-PRD-003-v1.md

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📅 PRD-007: SSO Integration
🚧 Status: In Progress | P1 | Day 3

Snippet: "...building on <mark>OAuth</mark> foundation from PRD-003, add SAML..."

📄 File: product/prds/03-in-progress/241020-sso-PRD-007-v1.md

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📅 PRD-012: Third-Party API Auth
✅ Status: Ready | P0

Snippet: "...reuse <mark>OAuth</mark> client credentials flow for API-to-API..."

📄 File: product/prds/02-ready/241023-api-auth-PRD-012-v1.md

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 Quick Actions:
  • View PRD-003: Read PRD-003 file
  • Copy decision: "Used OAuth 2.0 with PKCE flow"
  • Find more: /recall "authentication" --status=complete

Step 4: Semantic Search (Optional)

If embeddings are enabled, use vector similarity:

/recall --similar-to=PRD-003

→ Finding similar PRDs...
→ Using embeddings for semantic search...

Found 4 similar PRDs (by concept, not keywords):

1. PRD-007: SSO Integration (89% similar)
   Reason: Both involve authentication flows

2. PRD-015: Password Reset (76% similar)
   Reason: User identity verification

3. PRD-021: API Keys Management (68% similar)
   Reason: Authentication credentials

4. PRD-009: User Permissions (61% similar)
   Reason: Authorization system

Memory Index Structure

.claude/memory/
  ├── index.db                    # SQLite FTS5 full-text index
  ├── embeddings/                 # Vector embeddings (optional)
  │   ├── PRD-003.json
  │   ├── PRD-007.json
  │   └── index.json
  │
  ├── compressed/                 # Archived conversations (future)
  │   ├── 2025-10-01.jsonl.gz
  │   └── 2025-10-15.jsonl.gz
  │
  └── config.json                 # Memory settings

SQLite Schema

-- Main PRD content table
CREATE TABLE prds (
  prd_id TEXT PRIMARY KEY,
  name TEXT NOT NULL,
  status TEXT NOT NULL,
  priority TEXT,
  grade TEXT,
  file_path TEXT NOT NULL,
  content TEXT NOT NULL,
  created_at DATETIME DEFAULT CURRENT_TIMESTAMP,
  completed_at DATETIME,
  updated_at DATETIME DEFAULT CURRENT_TIMESTAMP
);

-- FTS5 full-text search table
CREATE VIRTUAL TABLE prds_fts USING fts5(
  prd_id UNINDEXED,
  name,
  content,
  decisions,
  blockers,
  learnings,
  tech_stack,
  tokenize = 'porter unicode61'
);

-- Trigger to keep FTS in sync
CREATE TRIGGER prds_ai AFTER INSERT ON prds BEGIN
  INSERT INTO prds_fts (prd_id, name, content, decisions, blockers, learnings, tech_stack)
  VALUES (new.prd_id, new.name, new.content, ...);
END;

-- Index for filters
CREATE INDEX idx_prds_status ON prds(status);
CREATE INDEX idx_prds_priority ON prds(priority);
CREATE INDEX idx_prds_completed_at ON prds(completed_at);

Embeddings (Optional)

For semantic search, generate embeddings using OpenAI API:

// Generate embedding for PRD
const embedding = await openai.embeddings.create({
  model: "text-embedding-3-small",
  input: prd.content
});

// Save to .claude/memory/embeddings/PRD-003.json
{
  "prd_id": "PRD-003",
  "embedding": [0.123, -0.456, ...], // 1536 dimensions
  "generated_at": "2025-10-26T10:00:00Z"
}

// Find similar PRDs using cosine similarity
function cosineSimilarity(a, b) {
  const dotProduct = a.reduce((sum, val, i) => sum + val * b[i], 0);
  cons