Cognee

Overview

Cognee is an open-source memory engine (27.6k⭐ on GitHub) with an ECL (Extract-Contrast-Link) pipeline and production-grade knowledge-graph infrastructure. It is the most advanced project in the vault’s ecosystem for backlinks and typed relationships at scale: its hosted “Karpathy Wiki” maintains 633 nodes and 1090 edges. It exposes a Cognee MCP server with 14 tools (remember/recall) and supports 38+ data sources.

Key Facts

• Stars: ~27.6k (one of the most-starred memory engines in the ecosystem)
• Architecture: ECL pipeline (Extract → Contrast → Link) over a knowledge graph (Neo4j / Postgres) + vector database
• MCP: cognee-mcp with 14 tools — the graph-backed MCP option in the comprehensive report’s Pattern C
• Scale: millions of documents, multi-source, multi-agent
• Backlinks/relations: computable knowledge graph (nodes + typed edges), not just visual wikilinks — this is what makes it lead in backlink/relation depth among documented projects
• Raised $7.5M seed (Pebblebed) in May 2026 — commercial validation of memory infrastructure
• Hosts “Karpathy Wiki” (karpathywiki.com): 633 nodes, 1090 edges

Features

Core Architecture

• ECL pipeline — Extract → Contrast → Link over knowledge graph
• Neo4j / Postgres + vector DB — graph + vector hybrid storage
• 27.6k stars — most-starred memory engine in ecosystem
• $7.5M seed — commercial validation (Pebblebed, May 2026)

Ingest & Processing

• 38+ data sources — documents, web, APIs, databases
• Multi-agent — supports multiple agent workflows
• Millions of documents — enterprise-scale ingestion
• ECL extraction — structured entity and relationship extraction

Search & Retrieval

• Hybrid search — graph traversal + vector similarity + keyword
• Knowledge graph queries — Cypher/SQL over Neo4j/Postgres
• Multi-source reasoning — links across 38+ sources via ECL

Knowledge Graph

• 633 nodes, 1090 edges — hosted Karpathy Wiki demonstrates scale
• Typed edges — explicit relationship types between entities
• Backlinks — computable, not just visual wikilinks
• Contrast step — identifies contradictions across sources

MCP / Tools

• cognee-mcp — 14 tools (remember/recall pattern)
• Graph-backed MCP — Pattern C in comprehensive report
• Agent integration — works with Claude, GPT, any MCP agent

Quality & Governance

• Contrast step — automatic contradiction detection
• Backlink validation — ensures graph integrity
• Production-grade — enterprise-ready infrastructure

Integration

• MCP protocol — 14-tool server for agent access
• Multi-agent — supports team workflows
• Karpathy Wiki — hosted demo at karpathywiki.com

Data & Storage

• Neo4j / Postgres — graph database backend
• Vector database — semantic search storage
• Knowledge graph — nodes + typed edges
• Markdown output — wiki pages in markdown

Relationships

• “implements” knowledge-graph — Cognee is the strongest production-grade knowledge-graph implementation documented here
• “belongs to” agent-memory-systems — belongs to the memory-first / graph-backed family
• “implements” llm-wiki — Pattern C (Graph-Backed) implements the LLM Wiki pattern on top of Cognee
• “exposes” model-context-protocol — exposes a 14-tool MCP server for agent access
• “references” rohitg00 — both extend the implicit wikilink graph toward typed/computable relations
• “differs from” rightmemory — sibling graph-backed memory substrate for coding agents

Implications

Cognee represents the high end of the ecosystem on the three dimensions you asked about: it has real backlinks (graph edges), typed relationships (knowledge graph), and generates new knowledge by linking across 38+ sources via ECL. For this vault, Cognee is the reference target if/when we move beyond implicit wikilinks to an explicit graph layer (the vault’s vault-roadmap and connection-methods note this transition happens past ~100 pages).

External Links

• Repository: https://github.com/topoteretes/cognee
• Karpathy Wiki (Cognee-hosted): https://github.com/topoteretes/karpathy-wiki
• Cognee MCP is referenced in the comprehensive report’s MCP integration patterns

Sources

^[raw/articles/llm-wiki-okf-comprehensive-report-2026.md]