cognee
cognee.ai
Memory for AI Agents in 5 lines of code
Description
Подключает источники знаний компании — Slack, GitHub, Linear, Notion, Google Drive, документы, чаты, тикеты, код и логи агентских запусков — в единый permission-aware граф, который могут запрашивать и агенты, и люди; сохраняет и извлекает воспоминания через SDK (cognee.remember, cognee.recall) с сохранением цитат и ссылок на источники; автоматически строит онтологии домена с сущностями, связями и правилами; принимает обратную связь и корректирует веса выдачи; индексирует репозитории кода; работает с Claude Code, Codex, MCP и другими агентными средами. На Enterprise-уровне добавляются би-темпоральная память с разрешением конфликтов, провенанс каждого ответа и персонализация по пользователю и агенту.
Cognee implements scalable, modular ECL (Extract, Cognify, Load) pipelines that allow you to interconnect and retrieve past conversations, documents, and audio transcriptions while reducing hallucinations, developer effort, and cost.
Installation
pip install cognee| Installs into | python, docker, mcp, cloud, self-host |
Not sure where to start — ask an assistant to walk you through:
Features
Modular: Cognee is modular by nature, using tasks grouped into pipelines
Local Setup: By default, LanceDB runs locally with NetworkX and OpenAI.
Vector Stores: Cognee supports LanceDB, Qdrant, PGVector and Weaviate for vector storage.
Language Models (LLMs): You can use either Anyscale or Ollama as your LLM provider.
Graph Stores: In addition to NetworkX, Neo4j is also supported for graph storage.
Use cases
Memory for AI Agents
Ontology definition
Entity resolution
Chatbot memory
FAQ
Cognee is an open-source AI memory engine that transforms unstructured raw data into a structured, persistent, and dynamic AI memory. It combines vector search and knowledge graphs to enable AI agents to understand, reason, and adapt through contextual memory storage and retrieval.
Cognee processes data through multiple key steps. These steps include Add, for ingesting raw data asynchronously; Cognify, for intelligently chunking documents, creating vector embeddings, extracting entities and relationships, and building a knowledge graph representing concepts and their interconnections; Search, for performing contextual queries that combine vector similarity and graph traversal to retrieve precise answers, often supported by large language models (LLMs); and Memify (coming soon), for further semantic enrichment of the knowledge graph for deeper context.
Unlike typical RAG assistants that rely mostly on keyword search and simple vector retrieval, Cognee builds a knowledge graph that maps conceptual relationships and dependencies across data. This allows it to answer complex, multi-hop queries across related entities, provide verifiable, context-grounded responses, and reduce hallucinations by grounding LLM output in fact-based memory stored externally.
The main components of Cognee’s memory system include a hybrid memory layer combining vector embeddings and graph database storage, persistent memory capable of learning and updating from user feedback and interactions, and flexible querying supporting semantic search, graph-based queries, and combined approaches.
Cognee acts as an external, verifiable memory layer that supplies LLMs with domain-specific, factual data rather than relying solely on the model’s internal probabilistic knowledge. This hybrid graph-vector memory supports precise retrieval of data relevant for a given query, significantly reducing hallucinations and improving trustworthiness.
Typical use cases or applications include building smart domain-specific copilots that learn and adapt over time, creating intelligent FAQ assistants that understand documentation context and relationships between concepts rather than just keyword matching, and enhancing AI agent memory in multi-agent frameworks or enterprise data environments.
Developers can get started with Cognee by ingesting and cleaning data, ideally removing noise like cookie banners and navigation for better graph quality. Then, use the .add() method to input data and run .cognify() to build the knowledge graph. Developers should perform searches with cognee.search() configured for appropriate query types such as graph completion, similarity, and insights. Cognee supports integrations with multiple vector and graph databases, as well as various LLM providers.
Cognee provides advanced features such as multi-type search including graph-based completion, semantic similarity, and combined insights, persistent memory of workflows and rules that can learn from interactions, and scalability from simple chatbots to complex multi-agent AI systems.
Yes, Cognee is open-source and available for developers to deploy, customize, and contribute.
Specs
| Type | Platform |
| Section | AI agents |
| Pricing | has a free tier (от $25/mo) |
| Platform | Desktop |
| Systems | macos, cli, linux, docker |
| Hosting | hybrid |
| Install | package |
| Installs into | python, docker, mcp, cloud, self-host |
| Who for | Инженеры и AI-команды, строящие агентов с долговременной памятью, компании, которым нужен общий «мозг» для агентов и сотрудников, пользователи кодовых агентов |
| Protocols | mcp |
| Site language | en |
| Vendor | Cognee (topoteretes) |
| GitHub | topoteretes/cognee |
| Rating | 0.00 (0 reviews) |
| Views | 538 |
| Launched | 2025-02-08 |
