Cognee

Cognee is an open-source AI memory engine built by Topoteretes UG, founded in 2024 by Vasilije Markovic and Boris Arzentar. The project started as an open-source initiative and has grown into a production-grade framework adopted by over 70 organizations, running more than one million pipelines per month. Cognee gives AI agents a shared, continuously improving memory by combining graph traversal and vector similarity into a single retrieval layer that goes well beyond standard RAG.

---

Core Products

Cognee Memory Engine

Cognee ingests data in any format and builds a structured knowledge graph with embeddings, making it queryable through four operations: remember, recall, forget, and improve. The default retrieval mode (GRAPH_COMPLETION) uses vector search as a hint to locate relevant graph triplets, then traverses the graph to assemble structured context before generating an answer. Memory graphs can be scoped per user, per group, or as a shared public graph, giving teams fine-grained multi-tenancy control.

Cognee UI

Starting with v0.3.3, Cognee ships a local web interface with interactive notebooks and a Graph Explorer. Developers can run cognee methods, build pipelines, and visualize query results directly in the browser without leaving their local environment.

---

Developer Resources

Cognee provides a Python SDK that wraps the full memory pipeline in three core operations: .add() to ingest data, .cognify() to build the knowledge graph, and .search() to retrieve context. The SDK supports text files, structured datasets, and custom data formats.

Helpful Links

• Documentation: getting started guide, API reference, and examples
• GitHub: source code, issues, and community contributions
• PyPI: install via pip install cognee
• LangChain integration: drop-in integration for LangChain projects

---

Key Features

Unified three-layer storage Cognee combines a graph store (Kuzu by default, also supports Neo4j, FalkorDB, Amazon Neptune, Memgraph), a vector store (LanceDB by default, also supports Qdrant, pgvector, Redis, Pinecone, ChromaDB), and a relational store (SQLite by default, also supports PostgreSQL) into a single memory interface.

14 retrieval modes The framework ships with 14 configurable retrieval strategies. GRAPH_COMPLETION is the default and produces more structured, relationship-aware context than cosine-similarity chunk retrieval.

Automatic ontology management Cognee continuously updates ontologies as new data is ingested, so the knowledge graph stays current without manual schema maintenance.

Multi-tenancy and graph isolation Memory graphs can be instantiated per user, per group, or as shared public graphs, making it suitable for multi-user agent applications.

---

Use Cases

Agentic workflows with persistent memory Teams building AI agents that need to recall previous decisions, user preferences, or domain knowledge across sessions use Cognee to maintain a durable, queryable memory layer.

GraphRAG over private data Organizations with large internal document libraries (legal, medical, research) use Cognee to transform document collections into navigable knowledge graphs, improving retrieval accuracy over standard vector search.

Multi-agent coordination Shared graph instances let multiple agents read and write to the same memory space, enabling coordination without duplicating context across agent boundaries.