Cognee

Persistent memory for AI agents is the problem Cognee tackles, replacing ad hoc RAG stacks with pipelines that turn raw documents, conversations, and files into a queryable knowledge graph paired with vector embeddings. Data moves through modular tasks: content is chunked, entities and relationships are extracted with an LLM, concepts and ontologies are induced, and everything lands in pluggable backends, with support for multiple graph databases, vector stores, and relational stores rather than a fixed stack. Retrieval then combines graph traversal with semantic search, which improves multi-hop questions where plain vector lookup fails. The Python library installs with pip, runs fully self-hosted, and can use local LLMs and embedding models, and integrations exist for agent frameworks including CrewAI, LangGraph, Google ADK, and Claude Code via MCP. Licensed Apache-2.0 with a hosted platform as the paid tier, the project has grown to about 29,500 GitHub stars, and teams adopt it to give assistants durable memory across sessions instead of stateless context windows.