Cognee Documentation — Open-Source AI Memory Platform for Agents

Cognee is an Apache-2.0 memory layer that sits between your data and an LLM, replacing similarity-only retrieval with a self-hosted knowledge graph built from your own documents. The documentation opens with Getting Started — install via pip install cognee or uv pip install cognee, plus a quickstart — and then a Core Concepts section that is the part worth reading in full. Its Architecture page describes the three stores Cognee keeps in sync: relational storage for documents and provenance, vector storage for semantic similarity, and graph storage for entities and relationships. A Building Blocks page covers DataPoints, the structured units that become graph nodes, along with Tasks and Pipelines. A Main Operations page covers the high-level Remember, Recall, Improve and Forget verbs alongside the lower-level Add, Cognify, Search and Memify calls that give granular pipeline control. Further Concepts covers node sets, the agent-memory decorator, RDF/XML ontologies, loaders and chunkers. Setup and Configuration documents pluggable backends: LanceDB as the default vector store with PGVector, Turso and community Qdrant, Chroma, Weaviate and Milvus adapters, and Kuzu as the default graph store with Neo4j, Postgres, Turso and Neptune alternatives. Separate sections cover the Cognee MCP server for Cursor and Claude Code, the managed Cognee Cloud, and REST and Python API references. The repository carries 30.4k stars and roughly 9,800 commits.

Topic

Agentic

Level

Intermediate

Format

Documentation

Cost

Free

Duration

Self-paced reference; ~3-4 hours to work through Getting Started and Core Concepts with the quickstart running

Provider

Cognee (topoteretes)

Hands-on

Yes — code/exercises

Certificate

None

• ✓Model documents as DataPoints that become typed nodes in a knowledge graph
• ✓Run the cognify pipeline: chunking, entity extraction, concept and ontology induction
• ✓Choose between the Remember/Recall/Improve/Forget verbs and the lower-level Add/Cognify/Search/Memify calls
• ✓Swap vector and graph backends — LanceDB, PGVector, Neo4j, Kuzu, Neptune — through configuration
• ✓Ground a graph in an external RDF/XML ontology instead of ad-hoc entity labels
• ✓Expose Cognee as an MCP server so Cursor or Claude Code can read agent memory
• ✓Compose custom Tasks into Pipelines rather than accepting one fixed ingestion flow

• •Graph-first rather than vector-first: retrieval follows entity relationships, so relational and causal questions become answerable instead of only 'find similar text'
• •Genuinely pluggable — several graph and vector backends are documented, so it fits an existing Postgres, Neo4j or LanceDB estate rather than demanding a new one
• •Reports 0.79 on the BEAM long-context benchmark at 100K tokens against a 0.735 prior state of the art, and 0.67 at 10M tokens against 0.641
• •30.4k GitHub stars, Apache-2.0 and roughly 9,800 commits — mature enough to self-host rather than a weekend project
• •Independent write-ups are candid about the costs: entity disambiguation needs multiple LLM passes and dedup logic, incremental updates are expensive enough that full re-ingestion is often preferred, and TypeScript support lags Python

This page was written from 4 sources, 2 on domains other than docs.cognee.ai.

1. 1.docs.cognee.ai — docs.cognee.aivendor
2. 2.docs.cognee.ai — introduction.mdvendor
3. 3.github.com — cognee
4. 4.memgraph.com — from rag to graphs cognee ai memory