에이전트 및 멀티 에이전트

Cognee: Open-Source Persistent Memory for AI Agents

topoteretes/cognee

Cognee is a Python-based AI memory platform that turns ingested data into persistent, self-hosted knowledge graph memory. It combines vector embeddings, graph reasoning and ontology generation so agents can retrieve context across sessions.

★ 29.7K별점
⑂ 2.9K포크 수
528미해결 이슈
Python언어
Apache-2.0라이선스
Q@project.QualityScore편집 점수

개요

Cognee provides a memory layer for AI agents through four primary operations: remember, recall, forget and improve. It supports permanent graph-backed memory and faster session memory that synchronizes to the graph. The repository also describes CLI, API server, MCP server, local UI, Docker, managed cloud and remote-server workflows. Repository metadata reports 29,718 stars, 2,872 forks and 528 open issues.

주요 기능

  • Persistent long-term memory across agent sessions
  • Self-hosted knowledge graph construction
  • Vector embeddings combined with graph reasoning
  • Cognitive-science-grounded ontology generation
  • Permanent and session-scoped memory
  • Remember, recall, forget and improve operations
  • Automatic routing among recall strategies
  • Python API, CLI, API server and MCP server
  • Local UI and Docker deployment options
  • Claude Code and OpenClaw plugin availability
  • Official Rust and TypeScript clients
  • Local embedded development with SQLite, LanceDB and Kuzudb
  • Optional Postgres, pgvector, Neo4j, Neptune and Redis backends
  • Community adapters for Qdrant, ChromaDB, Weaviate and Milvus
  • User or tenant isolation, traceability, audit traits and OTEL collector support are described in the README

요구 사항, 설치 및 빠른 시작

For Python, install with `uv pip install cognee`; the README also states that pip, Poetry or another preferred Python package manager can be used. Configure `LLM_API_KEY` in the environment or copy the supplied `.env.template`. A Postgres-enabled installation is shown as `pip install "cognee[postgres]"`. Docker Compose and prebuilt `cognee/cognee:main` and `cognee/cognee-mcp:main` images are also documented.

사용 정보

Minimal Python flow: import `cognee`, set `LLM_API_KEY`, call `await cognee.remember("Cognee turns documents into AI memory.")`, then call `await cognee.recall("What does Cognee do?")`. A `session_id` can be supplied for session memory, and `await cognee.forget(dataset="main_dataset")` removes a dataset. CLI equivalents include `cognee-cli remember`, `cognee-cli recall` and `cognee-cli forget --all`. Run `cognee-cli -ui` to open the local UI when a compatible container runtime is available.

모델 호환성 및 사용 사례

The quick start uses an OpenAI API key as its example. The README states that other LLM providers can be integrated through provider configuration, but specific supported models and a complete provider list are not stated in the supplied repository data.

라이선스 및 위험 참고 사항

Repository metadata identifies the license as Apache-2.0. The README warns that using Postgres as a graph store is a demo feature and states that the production-ready Postgres graph capability is available as a licensed product.

릴리스 및 유지 관리

No tagged release version or release changelog is included in the supplied repository record. The default branch is main, the repository is not archived, and the supplied metadata reports a last-pushed timestamp of 2026-08-03T09:10:45Z.

Firecrawl

firecrawl/firecrawl

★ 161.1KTypeScript

LangChain

langchain-ai/langchain

★ 143.6KPython

RAGFlow

infiniflow/ragflow

★ 86.7KGo