Cognee

Cognee ingests unstructured data (documents, chat history, code) and builds a knowledge graph plus vector index that AI agents can query for context across sessions. The pitch against plain RAG is that chunk-based vector search loses the relationships between facts, so multi-hop questions ("which of these three vendors reported an incident, and when") fail; a graph structure preserves those links. It ships as a self-hosted Python library/pipeline (github.com/topoteretes/cognee) and as a hosted "Cognee Cloud" service with the same core engine.

The clearest way to place Cognee is against Mem0, the other widely-used agent-memory library. Mem0 is built for single-agent, single-user memory (personalizing one user's chat history) and gates its knowledge-graph feature behind a paid tier. Cognee is built for multi-agent memory — several agents reading and writing the same graph — and includes graph capability in the free/open-source tier from the start. Third-party technical writeups (e.g. a comparison at vectorize.io) generally agree with this framing: graph-based recall wins on multi-hop correctness, vector-only recall wins on raw simplicity and latency for single-fact lookups. Benchmark numbers Cognee cites for itself (e.g. its BEAM-benchmark results) come from its own blog, not an independent source, and should be read as a vendor claim.

The core repo is Apache-2.0 with a single, unambiguous LICENSE file (no split code/docs licensing). It's substantially active: 30,129 GitHub stars, 2,938 forks, 359 open issues, a push to main within the last 24 hours, and a PyPI release (v1.5.0) published the same week this was checked. Releases are frequent — four tagged releases in the two weeks before this check, including active dev-preview tags — which is a healthy sign for a tool you'd depend on in production, not a healthy sign in itself of code quality.

Self-hosting the open-source engine is free indefinitely — you run it on your own infrastructure and pay only for whatever LLM/vector-store calls it makes. The managed Cognee Cloud has three tiers, all with real numbers published (not "contact us" placeholders): Free ($0/month, 1M tokens included, 1 workspace, unlimited users and API calls), Standard ($2.50 per 1M tokens processed, plus $5 per additional workspace, adds Slack/Notion/Google Drive source integrations and in-app support), and Enterprise (custom pricing, adds a dedicated support engineer, bring-your-own-cloud, and SLAs). At meaningful token volume the Standard tier's usage-based pricing is the number to model against your expected ingestion volume, not the flat access-fee model most SaaS tools use.

Cognee is built by a Berlin-based team (legal entity trades as topoteretes/Cognee) that raised a $7.5M seed round led by Pebblebed, with 42CAP and Vermilion Ventures participating, announced February 2026 and reported independently by EU-Startups and Pulse2 in addition to the company's own post. The vendor's /trust page states GDPR-aligned processes maintained via a third party (heyData), a DPA available on request, and three deployment options including fully local ("your data never leaves your infrastructure") — but it does not claim SOC 2 or any other formal security certification, and there's no separate named security/compliance audit page beyond that one trust page.

Cognee has zero reviews on G2 and none found on Trustpilot as of this check — it's simply too early for independent user review volume to exist yet, which is normal for a tool that hit 1.0 in June 2026, not a red flag on its own. The customer names the vendor lists (Bayer, University of Wyoming, Knowunity, SlideSpeak, Dynamo, DeepMetis) are vendor-published; the University of Wyoming one has a detailed named case study on Cognee's own blog, but none of the six were independently corroborated in this research. Treat vendor figures like "5M+ SDK runs/month" and "70+ companies running it live" the same way — plausible given the GitHub activity, but not independently verifiable.