cognee-mcp
Officialby topoteretes
The cognee MCP server connects Claude Code and other MCP clients to cognee's knowledge graph memory engine. Once connected, Claude gains persistent memory tools — it can ingest documents and code into a structured knowledge graph, query that memory using graph+vector hybrid search, analyze entire code repositories, and track developer rules across sessions. This transforms Claude from a stateless assistant into an agent with genuine persistent memory that survives across conversations.
databaseai-memoryknowledge-graphragpersistent-memoryclaudeagentspythonvector-search
Installation
npx cognee-mcp
Key Features
- ✓Persistent cross-session memory — Claude stores and retrieves knowledge that persists beyond a single conversation window
- ✓Code repository analysis — the codify tool builds a full code graph from a repository for architecture-level questions
- ✓Multi-mode search — seven distinct search modes including GRAPH_COMPLETION, RAG_COMPLETION, CODE, CHUNKS, SUMMARIES, CYPHER, and FEELING_LUCKY
- ✓Developer rules bootstrap — automatically ingests .cursorrules and similar files so Claude consistently applies project conventions
- ✓Background processing — cognify and codify run as background pipelines with status tracking, so Claude doesn't block during large dataset processing
- ✓Multiple transports — SSE, Streamable HTTP, and stdio support for Claude Desktop, Claude Code, Cursor, Cline, and Continue
Use Cases
- →Long-running coding projects where Claude needs to remember codebase architecture, past decisions, and developer rules across sessions
- →Internal knowledge base Q&A — ingest company docs into cognee once; Claude queries them with relationship-aware retrieval in any conversation
- →Multi-session research where Claude accumulates and connects findings over days, building a growing knowledge graph
- →Personalized AI assistant that improves from past Q&A interactions logged via save_interaction
- →Code review with full context — use codify to give Claude deep understanding of module dependencies before reviewing PRs