Cognee bets agents need memory, not bigger context — 27.6k stars and a $7.5M seed

Cognee is an open-source AI memory platform: it ingests data in any format and continuously builds a self-hosted knowledge graph plus vector store so agents keep persistent long-term memory across sessions. The Apache-2.0 Python core is free, with paid Cognee Cloud for managed memory and plugins for Claude Code and OpenClaw.

That agents' real bottleneck is memory, not context size: open-source a graph-plus-vector memory layer, win developers, then charge for the managed cloud.

$7.5M seed on February 19, 2026, led by Pebblebed with 42CAP and angels including people from Google DeepMind and n8n (startup.eu; NewDecoded); prior pre-seed amounts are undisclosed.

The Cognee team — with backgrounds in economics, clinical psychology, large-scale production engineering and graph theory — started from the conviction that LLMs cannot remember context across sessions and that plain RAG frameworks had too many knobs and no state. They open-sourced the core under Apache 2.0 and launched on Hacker News on June 3, 2025 with a Show HN describing a memory layer that 'remembers context', together with a research paper (arXiv:2505.24478).

Adoption climbed through the agent wave: by February 2026 the project passed 12,000 GitHub stars and one million pipeline runs, with Bayer and the University of Wyoming named as users, and Cognee closed a $7.5M seed led by Pebblebed. The money was earmarked for a managed cloud platform, a Rust engine for edge devices and local privacy, and research translating human-memory models into production software; the repo kept growing to 27.6k stars by mid-July 2026.

Still live and scaling: 27.6k GitHub stars on the July 13, 2026 trending list, seed capital banked, and a paid cloud product in development alongside integrations for Claude Code and OpenClaw. Revenue and customer counts beyond the disclosed 70+ companies have not been published.

Context windows were growing faster than agents could use them, while statelessness caused hallucinations — making persistent memory the visible bottleneck the team could own.

RAG frameworks required too many tuned knobs and held no state, leaving room for a product that combined graph structure, vectors and rules under one API.

An Apache-2.0 core turned every agent developer into a potential user, and the 2026 agent boom supplied the adoption curve.

Memory is a horizontal layer: any agent framework (Claude Code, OpenClaw, custom stacks) can plug in, so one open-source project could ride many platforms at once.

The managed cloud and enterprise permissions gave a path from free developers to paying companies without open-core trapdoors that break community trust.

In an agent gold rush, sell the layer every agent needs rather than another agent: the open-source core gave Cognee distribution; the seed funded the cloud that turns developers into revenue.

As of September 2026 Cognee is live and scaling: the repo passed 27.6k GitHub stars by July 13, 2026, the $7.5M seed from Pebblebed and 42CAP is banked, and work is aimed at Cognee Cloud plus a Rust engine for local and edge deployments. Its disclosed traction — one million+ pipeline runs, 70+ companies, and named users Bayer and the University of Wyoming — dates to the February 2026 seed announcement; no newer revenue or customer figures have been published.