<- all tokdocs
2026.09.16@howtowebdev

Cognee: Open-Source Memory Layer That Gives AI Agents a Self-Hosted Knowledge Graph

AI
Watch on TikTok

View on TikTok ->

Cognee is an Apache-2.0 Python library from Berlin startup Topoteretes that turns documents, code, and conversations into a knowledge graph AI agents can query across sessions. The video is a 32-second scroll through the GitHub README and docs site with a voiceover reading the pitch. The transcript renders the name as "Cogni," but every frame shows the repo at topoteretes/cognee, and the caption links there directly.

What the video shows

The frames are a screen recording of two pages. The first is the GitHub repo, where the README badges read 30k stars, 3k forks, 10.3K commits, latest tag v1.5.4, 1M downloads, Apache-2.0 license, and 275 contributors, plus a "GitHub Trending #1 Repository Of The Day" badge. The folder listing includes cognee-mcp, cognee-frontend, cognee-starter-kit, and a .claude/skills directory. The second page is docs.cognee.ai, headed "Build AI Memory with Cognee," with a diagram labeled Remember, Improve, Recall.

The voiceover describes cognee as "an open-source AI memory platform and knowledge graph engine for autonomous agents" that "ingests code, documentation, and chat logs to construct a self-hosted knowledge graph." That wording is lifted almost verbatim from the repo's description line.

What cognee actually is

The docs intro shown in the frames is the clearest definition: cognee is a memory layer for AI agents. It is not a model. It calls whatever LLM you already use (Claude, GPT, a local model) to build and query memory, then hands the right context back to that model on each call. The docs contrast it with plain RAG: instead of embedding chunks and retrieving by similarity, cognee adds a graph of entities and relationships on top, so recall can follow connections rather than match text.

Three operations organize the API: remember (ingest and store), improve (enrich or promote session memory into the permanent graph via .improve), and recall (query). Session memory is the fast short-term path; the permanent graph runs the full ingestion and graph-building pipeline.

Verified as of this writing: the GitHub repo reports 30.8k stars and Apache-2.0. PyPI lists version 1.5.4, released September 4, 2026, for Python 3.10 through 3.14. Install is pip install cognee or uv pip install cognee.

Who makes it

The company site lists the legal entity as Topoteretes UG in Berlin, with Vasilije Markovic as managing director. The README links a 2025 arXiv paper, "Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning" (Markovic, Obradovic, Hajdu, Pavlovic), which uses cognee as the test framework for tuning chunking, graph construction, retrieval, and prompting on multi-hop QA benchmarks. The company sells a managed Cognee Cloud and an enterprise tier alongside the open-source SDK.

How you plug it in

The docs list four entry points: Python, a CLI, a REST API, and an MCP server. The MCP server is how Cursor, Cline, and similar clients connect. Claude Code gets a direct integration via a marketplace plugin, which explains the .claude/skills folder in the repo root. The README also mentions TypeScript and Rust SDKs, which the video does not cover.

Claims to treat with caution

The video's "perfect for AI engineers and enterprise developers" line is marketing, not a finding. Two other items could not be pinned down:

  • The "#1 Repository Of The Day" badge is a static image in the README. I could not confirm which day it refers to.
  • The "1M downloads" badge has no visible source. The company site separately claims "5M+ monthly SDK runs." Neither figure is independently verified here.

The README's "ingest data in any format" is broader than the video's "code, documentation, and chat logs." Treat the video's list as examples, not limits.

Key Takeaways

  • Cognee is a memory layer, not a model. It sits between your data and whatever LLM you already run.
  • The core difference from RAG is a knowledge graph on top of the vector store, so retrieval follows entity relationships.
  • It is Apache-2.0, on PyPI at v1.5.4, and requires Python 3.10 or newer.
  • Integration paths are Python, CLI, REST, MCP server, and a Claude Code plugin.
  • The maker is Topoteretes UG in Berlin, which also sells a hosted cloud version.
  • The transcript's "Cogni" is a transcription error. The project is cognee.

Resources

Published September 16, 2026. Writeup generated from a favorited TikTok.