topoteretes/cognee

📄 Read the research paper: Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning — Markovic et al., 2025

When to use Cognee

• Build a Company Brain. Bring documentation, conversations, tickets, code, and agent work into shared memory. Help your team and agents connect a decision to the discussion and implementation behind it. Explore Company Brain.
• Give agents memory across runs. Retain project context, past decisions, fixes, and learned rules. Distill useful session lessons into durable knowledge that another session can retrieve. Connect your agent.
• Ground agents in your domain. Structure memory around the entities and relationships your application needs, with custom data models and ontologies. Explore ontologies.

Choose your starting point

Quickstart

Requires Python 3.10–3.14.

1. Install Cognee with pip, uv, or your preferred Python package manager. The gliner extra brings the local extraction model used when no LLM key is configured:

uv pip install "cognee[gliner]"

Run locally without an LLM

2. Build and query memory. With no LLM key configured, Cognee extracts the graph with the local GLiNER model and embeds with a local embedding model; both download on first use.

Save this as quickstart.py and run python quickstart.py if you are feeling old school, or tell your LLM to do it:

import asyncio

import cognee


async def main():
    # Extract a knowledge graph and embed the text with local models.
    await cognee.remember(
        "Marie Curie was born in Warsaw and worked at the University of Paris.",
        dataset_name="local_quickstart",
    )

    # Retrieve the matching source text; no LLM generates an answer.
    results = await cognee.recall(
        "Where was Marie Curie born?",
        datasets=["local_quickstart"],
    )
    for result in results:
        print(result)


if __name__ == "__main__":
    asyncio.run(main())

The same workflow is available from the CLI:

cognee-cli remember "Marie Curie was born in Warsaw." -d local_quickstart
cognee-cli recall "Where was Marie Curie born?" -d local_quickstart

Text ingestion, retrieval, and session storage work without an LLM. LLM-dependent improvement stages skip automatically.

Generated answers and media processing that requires a vision or transcription model need additional LLM configuration.

The bundled GLiNER extractor is a demo of Cognee's small-model pipeline. For a production-ready version with higher accuracy and broader label coverage, reach out to us.

Optional: Configure the LLM

3. Add an LLM to get generated answers instead of retrieved passages:

import os

os.environ["LLM_API_KEY"] = "YOUR OPENAI_API_KEY"

Alternatively, create a .env file using our template.

Once a key is set, Cognee uses OpenAI for language models and embeddings, and processing and generated answers make provider calls. See installation, other providers, or local Ollama models for other setups.

Explore the bundled demo

To explore a prebuilt graph without downloading extraction or embedding models:

cognee-cli demo

This command works with the base pip install cognee package. It loads bundled sample data and runs keyword search without an API key. Use the local quickstart above to build a graph from your own text.

How Cognee works

Cognee builds connected memory from different sources. Text becomes entities, relationships, and searchable chunks; code becomes a graph of symbols and dependencies. Session distillation curates accepted lessons into permanent memory.

At query time, retrieval selects relevant graph, vector, or code context. Your application can inspect the retrieved evidence and use it to answer a question or continue an agent task.

Explore the architecture and session lifecycle.

Connect your agent

Install the Claude Code plugin:

claude plugin marketplace add topoteretes/cognee-integrations
claude plugin install cognee-memory@cognee

Or install the Codex plugin. Enable hooks first, either with the CLI:

codex features enable hooks

or in ~/.codex/config.toml:

[features]
hooks = true

Then add the marketplace and the plugin:

codex plugin marketplace add topoteretes/cognee-integrations --ref main
codex plugin add cognee@cognee

Follow the plugin setup guide to configure local or remote memory.

Browse the integrations repository for agent frameworks, plugins, and source connectors. Each guide describes its setup and memory capture behavior.

To inspect a local installation in the UI:

cognee-cli -ui

The UI launcher requires Node.js/npm; Docker is needed for its MCP service. See local UI setup.

Explore examples

• Build a small Company Brain from text, code, and session lessons.
• Build a Company Brain from your database, ticket export, and docs (or a sample company), then browse it in the UI and query it from Claude Code or Codex.
• Import memory from Mem0, Letta, Zep, or Graphiti using the COGX exchange format.
• Run with local Ollama models, including a local embedding model.
• Visualize your knowledge graph and inspect its connections.
• Browse runnable examples for ingestion, sessions, feedback, and custom pipelines.
• Run the prebuilt API with Docker Compose or use the deployment templates.
• Explore community adapters and add-ons.

Deploy Cognee

For a local API demo using a prebuilt image, follow the minimal Docker Compose guide. It includes a persistent-volume configuration and explains the single-user demo settings. Or run the image directly:

docker run --rm -it -p 8000:8000 \
  -e LLM_API_KEY="sk-..." \
  -e ENABLE_BACKEND_ACCESS_CONTROL=false \
  -v cognee_storage:/cognee-storage \
  cognee/cognee:main

ENABLE_BACKEND_ACCESS_CONTROL=false is the single-user/local posture — without it the API defaults to multi-tenant mode and every /api/v1 call requires an authenticated user. To keep authentication on instead, set DEFAULT_USER_PASSWORD to make the default account loginable (see the compose guide). Note that --rm discards container-local data on exit; the -v cognee_storage:/cognee-storage named volume keeps your memory across runs.

To run the API, UI, and MCP server from a source checkout, clone this repository, enter its directory, copy .env.template to .env, and configure your providers. Then run:

docker compose --profile ui --profile mcp up

The default ports are API 8000, UI 3000, and MCP 8001. For deployment beyond a local demo, configure authentication, persistent storage, and compatible backends using the permissions guide and deployment templates. Cognee Cloud provides the managed option.

The cognee/cognee image ships with the GLiNER runtime baked in, so text ingestion and retrieval work in Docker without an LLM key; the local extraction and embedding models download on first use. Set LLM_API_KEY (as above) when you want generated answers.

Run the Whole Memory Layer on Postgres

Graph memory traditionally means operating a stack — a graph database for relationships, a vector database for embeddings, Redis for sessions, and a relational database for metadata — all deployed, secured, and paid for before an agent remembers anything. Since cognee 1.0 you can run the entire memory layer on a single Postgres instance.

⚠️ Warning: Using Postgres as a graph store is currently released as a demo feature. The production-ready version is available as a licensed product. Use the demo to keep relational metadata, PGVector, and graph state working together in one Postgres service.

Benchmarks and research

The BEAM evaluation measures conversational memory using synthetic long-context conversations and an LLM judge. The reported runs use Cognee's memory components with benchmark-specific data formatting, prompts, and retrieval configuration.

The two settings use different conversations, ingestion models, and retrieval-selection procedures. Read the methodology, models, limitations, and reproduction instructions before comparing these scores with other systems. The report also documents the remaining reproduction gap for the distributed 10M ingestion.

For the research behind Cognee's graph/LLM interface, see Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning (Markovic et al., 2025).

Latest News

• v1.6.1 — Google Sync & Visualization (September 24, 2026): Google Drive and Gmail OAuth connectors bundled in the SDK, /visualize/json streamed in chunks for large graphs, and the GLiNER installer moved off the event loop with CPU torch installed on first use.
• Document external_metadata is stamped on every chunk and surfaced in hybrid retrieval. Breaking: dlt is now a core dependency — make sure it is installed if you manage dependencies by hand.
• v1.6.0 — Keyless workflows & pipeline reliability (September 18, 2026): build and search text memory with local models and no cloud LLM key; LLM-dependent improvement stages skip when no LLM is configured, and pipeline recovery preserves completed documents after crashes.

Community & Support

Contributing

We welcome contributions from the community! Your input helps make Cognee better for everyone. See CONTRIBUTING.md to get started.

Code of Conduct

We're committed to fostering an inclusive and respectful community. Read our Code of Conduct for guidelines.

Research & Citation

We recently published a research paper on optimizing knowledge graphs for LLM reasoning:

@misc{markovic2025optimizinginterfaceknowledgegraphs,
      title={Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning},
      author={Vasilije Markovic and Lazar Obradovic and Laszlo Hajdu and Jovan Pavlovic},
      year={2025},
      eprint={2505.24478},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2505.24478},
}