luum-home/luum-cognitive-os/cognee-integration

Purpose

Integrate Cognee as a knowledge graph memory layer that complements Engram. While Engram provides flat observation-based memory, Cognee adds relationship-aware knowledge graphs with semantic search, enabling agents to understand connections between concepts, decisions, and patterns.

Invocation

/cognee-setup — Initial configuration /cognee-add <source> — Add knowledge from a source (file, URL, text) /cognee-search <query> — Search the knowledge graph

Setup

Prerequisites

• Python 3.10+
• pip install cognee
• Graph database backend: Neo4j (recommended) or NetworkX (lightweight)
• Vector store: Qdrant, Weaviate, or pgvector

Configuration

Add to cognitive-os.yaml services section:

services:
  cognee:
    mode: on_demand
    idle_timeout_minutes: 15
    config:
      graph_backend: "${COGNEE_GRAPH_BACKEND:-networkx}"  # networkx | neo4j
      vector_store: "${COGNEE_VECTOR_STORE:-lancedb}"
      llm_provider: "${COGNEE_LLM_PROVIDER:-ollama}"
      mcp_enabled: true
      mcp_port: 8100

Environment Variables

COGNEE_GRAPH_BACKEND=networkx     # or neo4j
COGNEE_VECTOR_STORE=lancedb       # or qdrant, weaviate, pgvector
COGNEE_LLM_PROVIDER=ollama
COGNEE_LLM_MODEL=llama3.1:8b
COGNEE_LLM_ENDPOINT=http://localhost:11434/v1
COGNEE_LLM_API_KEY=ollama
COGNEE_EMBEDDING_PROVIDER=fastembed
COGNEE_EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2
COGNEE_EMBEDDING_DIMENSIONS=384

What to Do

Step 1: Add Knowledge

Process documents, code, or text through the ECL pipeline:

import cognee
# Add source
await cognee.add("path/to/document.md")
# Process (Extract → Cognify → Load)
await cognee.cognify()

Step 2: Search Knowledge Graph

Query relationships and semantic connections:

results = await cognee.search("How does the auth middleware handle JWT tokens?")
# Returns: nodes, relationships, and semantic matches

Step 3: MCP Server Integration

Cognee provides an MCP server that Claude Code agents can use directly:

# Start Cognee MCP server
cognee mcp serve --port 8100

Register in Claude Code settings for direct agent access.

Architecture: Engram + Cognee

Agent Query
    ├─ Engram (flat observations)
    │   └─ "What did we decide about X?" → topic-key lookup
    │
    └─ Cognee (knowledge graph)
        └─ "How does X relate to Y?" → graph traversal + semantic search

Rules

• Cognee complements Engram — it does NOT replace it
• Use Engram for: decisions, conventions, feedback, project state
• Use Cognee for: relationships, codebase understanding, knowledge synthesis
• Default to lightweight backend (NetworkX + LanceDB) for evaluation
• Upgrade to Neo4j + Qdrant only when graph size exceeds local capacity
• MCP server runs on-demand, not always-on (save resources)