300+ Tools CoveredSource Data Updated Weeklydates

Cognee

Cognee is an open-source memory layer for AI agents. It turns documents, code and conversations into a self-hosted knowledge graph that an agent writes with remember and queries with recall, with improve and forget for curation. For a buyer, the decisive facts are the licence, how it installs, and how much of it works before any model provider is involved.

Visit Cognee →
Type
Agent Memory
Category
AI Agents

By Egor Burlakov · Edited · Source dates below

Buying snapshot

Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.

Cognee at a glance

AttributeDetail

Licence

Apache-2.0, in the topoteretes/cognee repository.

Install

Python package via pip or uv; uv pip install "cognee[gliner]" adds the local extraction model used when no LLM key is configured. Requires Python 3.10–3.14.

Runs without an LLM key

Text ingestion, retrieval and session storage work with the local GLiNER extractor and a local embedding model, both downloaded on first use. LLM-dependent improvement stages skip automatically; generated answers and media that needs a vision or transcription model require LLM configuration.

Self-hosting

docker run cognee/cognee:main, or docker compose --profile ui --profile mcp from a source checkout. Default ports: API 8000, UI 3000, MCP 8001.

Access control default

Multi-tenant: every /api/v1 call requires an authenticated user unless ENABLE_BACKEND_ACCESS_CONTROL=false selects the single-user/local posture.

Managed option

Cognee Cloud is the managed option; the vendor also deploys Cognee BYOC inside a customer's own cloud.

Interfaces

Python API, TypeScript SDK, Cognee-RS for Rust, REST API, MCP server for Cursor and Cline, and memory plugins for Claude Code, Codex and OpenClaw.

Memory operations

remember stores content or code, in a session when a session ID is supplied; recall retrieves context and answers; improve applies feedback and bridges session knowledge into the graph; forget removes an item or dataset.

Single-Postgres memory layer

Since cognee 1.0 the whole memory layer can run on one Postgres instance, but Postgres as a graph store is released as a demo feature; the production-ready version is a licensed product.

Our take

Cognee is Apache-2.0 software you run yourself, so the cost basis is the stack you operate. A keyless install uses free small models on your own CPU, so the first evaluation costs machine time and storage. Once LLM_API_KEY is set, Cognee uses OpenAI for language models and embeddings, and processing and generated answers make provider calls — that usage, together with the Postgres or other backends you keep running, is what grows with volume.

If the question is whether graph memory helps an agent you already ship, start on the open-source path: the local quickstart and the MCP or Claude Code plugin let you test on your own data without a model key. If you need the production-ready extractor, or Postgres as a graph store in production, plan a vendor licence conversation into the timeline rather than expecting the bundled build to carry it.

Evaluate Cognee

Agent Memory landscape
2 published products in this market
Browse AI Agents
Every published tool in the category
Stack Recommender
Build a stack around Cognee

Where it stands out

  • A real evaluation can happen before procurement: building and querying text memory needs no model key and no cloud call, so your documents stay on the machine you chose during the trial.

  • Plugging into an existing agent is a packaged step rather than custom glue — marketplace installs for Claude Code and Codex, an MCP server for other clients, and SDKs in Python, TypeScript and Rust.

  • The four operations cover the whole memory lifecycle, including deletion of a specific item or dataset — useful when a buyer has to answer how memory is removed, not only how it is stored.

  • You keep placement control: run the prebuilt image with a named volume, add UI and MCP profiles from a checkout, or have the vendor deploy BYOC in your own cloud.

Trade-offs to weigh

  • What you evaluate is not what a production build would extract. The bundled GLiNER extractor is a demo of the small-model pipeline, and the vendor describes a production-ready version with higher accuracy and broader label coverage available by reaching out — so test extraction on your own documents and expect the recognised labels to change later.

  • A demo configuration is not a deployment. The API is multi-tenant unless you flip the access-control flag, --rm discards container-local data on exit, and persistence depends on mounting the cognee_storage volume — so plan authentication and storage deliberately.

  • The published BEAM results are narrow: 0.79 at a 100K-token context over four rounds of 20 questions from one held-out conversation, and an exploratory 0.67 at 10M tokens scored on the same question set across five rounds. The settings differ in conversations, ingestion models and retrieval selection, and the report documents a remaining reproduction gap for the distributed 10M ingestion — reproduce on your own data before leaning on either number.

  • Releases move fast and can break builds: v1.6.1 on 24 September 2026 made dlt a core dependency, flagged as breaking for hand-managed dependencies. Pin versions and read release notes before upgrading.

Public signals

About these signals

Verified factual signals from public sources. They indicate observable activity or interest, not total adoption, product quality, or cost.

2.8k GitHub commits 90d33.2k GitHub stars

See all signals from 2 sources
Source
Signals
Last updated
GitHub
Commits 90d:2.8kStars:33.2k
November 9, 2026
PyPI
Weekly downloads:42.8k
November 9, 2026