---

name: Knowledge version: 1.2.5 description: Personal knowledge management system using Graphiti knowledge graph with FalkorDB. USE WHEN user says 'store this', 'remember this', 'add to knowledge', 'search my knowledge', 'what do I know about', 'find in my knowledge base', 'organize my information', 'install knowledge', 'setup knowledge system', 'configure knowledge graph', or requests knowledge capture, retrieval, synthesis, installation, or configuration of their personal knowledge graph system.

Knowledge

Persistent personal knowledge system powered by Graphiti knowledge graph with FalkorDB backend. Automatically extracts entities, relationships, and temporal context from conversations, documents, and ideas.

Workflow Routing

Core Capabilities

Knowledge Graph Features:

• Automatic Entity Extraction - Identifies people, organizations, locations, concepts, preferences, requirements
• Relationship Mapping - Tracks how entities connect with temporal context
• Semantic Search - Finds relevant knowledge using vector embeddings
• Episode-Based Storage - Preserves context and conversations over time
• Multi-Source Input - Accepts text, JSON, messages, and structured data

Built-in Entity Types:

• Preferences - User choices, opinions, configurations
• Requirements - Features, needs, specifications
• Procedures - SOPs, workflows, how-to guides
• Locations - Physical or virtual places
• Events - Time-bound occurrences, experiences
• Organizations - Companies, institutions, groups
• Documents - Articles, reports, books, content

Prerequisites

Required Setup:

The pack is installed at ~/.claude/Packs/madeinoz-knowledge-system/ (or $PAI_DIR/Packs/madeinoz-knowledge-system/).

1. Start the Graphiti MCP server:

   Copycd ~/.claude/Packs/madeinoz-knowledge-system
   bun run src/server/run.ts
   

2. Verify server is running:

   Copycd ~/.claude/Packs/madeinoz-knowledge-system && bun run src/server/knowledge.ts health
   

3. Configure API key (in PAI .env ~/.claude/.env):

   CopyMADEINOZ_KNOWLEDGE_OPENAI_API_KEY=sk-your-key-here
   

Knowledge CLI (Preferred Interface)

The Knowledge CLI provides token-efficient, human-readable output. Always use the Knowledge CLI instead of direct MCP tool calls for better readability and reduced token consumption.

Run commands from the pack directory:

Copycd ~/.claude/Packs/madeinoz-knowledge-system

Commands:

Copy# Add knowledge
bun run src/server/knowledge.ts add_episode "Title" "Body" "Source"

# Search entities (30%+ token savings)
bun run src/server/knowledge.ts search_nodes "query" 10

# Search relationships (30%+ token savings)
bun run src/server/knowledge.ts search_facts "query" 10

# Get recent episodes (25%+ token savings)
bun run src/server/knowledge.ts get_episodes 10

# Get system status
bun run src/server/knowledge.ts get_status

# Clear graph (destructive - requires --force)
bun run src/server/knowledge.ts clear_graph --force

# Check server health
bun run src/server/knowledge.ts health

Options:

• --raw - Output raw JSON instead of compact format
• --metrics - Display token metrics after each operation
• --metrics-file <path> - Append metrics to JSONL file

What Gets Captured:

• Conversations and insights from work sessions
• Research findings and web content
• Code snippets and technical decisions
• Project documentation and notes
• Personal preferences and decisions
• Meeting notes and action items

Examples

Example 1: Capture a Learning

User: "Remember that when using Podman volumes, you should always mount to /container/path not host/path"

→ Invokes CaptureEpisode workflow → Stores episode with extracted entities:

• Entity: "Podman volumes" (Topic)
• Entity: "volume mounting" (Procedure)
• Fact: "Podman volumes use /container/path syntax" → User receives: "✓ Captured: Podman volume mounting syntax"

Example 2: Search Knowledge

User: "What do I know about Graphiti?"

→ Invokes SearchKnowledge workflow → Searches knowledge graph for "Graphiti" entities → Returns related entities, facts, and summaries → User receives: "Based on your knowledge graph, Graphiti is..."

Example 3: Find Relationships

User: "How are FalkorDB and Graphiti connected?"

→ Invokes SearchFacts workflow → Searches for edges between FalkorDB and Graphiti entities → Returns facts showing relationship with temporal context → User receives: "FalkorDB is the graph database backend for Graphiti MCP server"

Example 4: Get Recent Learning

User: "What did I learn this week about PAI?"

→ Invokes GetRecent workflow → Retrieves recent episodes mentioning "PAI" or "Personal AI Infrastructure" → Returns chronological list with timestamps → User receives: "Recent additions: 1) PAI skills architecture... 2) Canonical skill structure..."

Example 5: Clear and Reset

User: "Clear my knowledge graph and start fresh"

→ Invokes ClearGraph workflow → Confirms destructive action → Deletes all entities and relationships → Rebuilds indices → User receives: "✓ Knowledge graph cleared. Ready for fresh knowledge capture."

MCP Integration

MCP Server Endpoint:

Copyhttp://localhost:8000/mcp/

Available MCP Tools:

Naming Convention (Hybrid Approach):

• User-facing (Skills/Workflows): Knowledge-friendly language ("store knowledge", "search my knowledge")
• Internal (TypeScript): Graphiti-native methods (addEpisode, searchNodes, searchFacts)
• MCP Layer: Actual tool names (add_memory, search_memory_nodes, search_memory_facts)

Response Caching: Search operations (search_memory_nodes, search_memory_facts) are cached to improve performance:

• TTL: 5 minutes (configurable via cacheTtlMs)
• Max entries: 100 (configurable via cacheMaxSize)
• Scope: Per-client instance (not shared across sessions)
• Cache invalidation: Automatic on TTL expiry, or manual via clearCache()

To disable caching, initialize the client with enableCache: false.

Configuration Options

Environment Variables (set in PAI config: $PAI_DIR/.env or ~/.claude/.env):

Copy# LLM Configuration
MADEINOZ_KNOWLEDGE_OPENAI_API_KEY=sk-your-key-here
MADEINOZ_KNOWLEDGE_MODEL_NAME=gpt-4o-mini
MADEINOZ_KNOWLEDGE_LLM_PROVIDER=openai
MADEINOZ_KNOWLEDGE_EMBEDDER_PROVIDER=openai

# Concurrency (adjust based on API tier)
MADEINOZ_KNOWLEDGE_SEMAPHORE_LIMIT=10

# Group ID (for multiple knowledge graphs)
MADEINOZ_KNOWLEDGE_GROUP_ID=main

# Disable telemetry
MADEINOZ_KNOWLEDGE_GRAPHITI_TELEMETRY_ENABLED=false

Model Recommendations:

• gpt-4o-mini - Fast, cost-effective for daily use
• gpt-4o - Better for complex reasoning
• gpt-3.5-turbo - Economy option, may miss some entities

Related Documentation

• ${PAI_DIR}/skills/CORE/SkillSystem.md - Canonical skill structure guide
• ${PAI_DIR}/skills/CORE/SYSTEM/MEMORYSYSTEM.md - PAI's memory documentation
• Graphiti Documentation
• Podman Configuration

Last Updated: 2025-01-03