The archive · AI & Models · Product decision · 2024–2026
Cognee bets agents need memory, not bigger context — 27.6k stars and a $7.5M seed
Berlin's Cognee builds an open-source graph-plus-vector memory layer for AI agents, hit 27.6k GitHub stars by July 2026, and raised a $7.5M seed in February.
Cognee
What the business is
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.
Starting capital：$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.
How it started
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).
What happened
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.
How it ended up
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.
Background
Cognee is an open-source AI memory platform from Berlin: it ingests documents, messages, databases or code and continuously builds a self-hosted knowledge graph plus vector store, giving AI agents persistent long-term memory across sessions. The team's founding argument was that LLMs are stateless — they cannot remember context, which causes hallucinations and forces developers to stuff unmanaged context into every prompt.
Its wedge was architectural: instead of another vector-only RAG framework, Cognee combines graph, vector and relational stores through an ECL pipeline (Extract, Cognify, Load), so retrieval can follow relationships between facts rather than only embedding similarity, with developer-defined ontologies and rules. The core shipped under Apache 2.0, and the founders launched on Hacker News in June 2025 alongside a research paper.
Adoption rode the agent wave: 12,000+ GitHub stars and one million+ pipeline runs by February 2026, users including Bayer and the University of Wyoming, and a $7.5M seed led by Pebblebed to build a managed cloud and an edge-focused Rust engine. By July 13, 2026 the repo sat at 27.6k stars on Refft's monthly GitHub-trending list, making Cognee one of the fastest-growing open-source startups of the period.
What has to be true
- 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.
What can be applied
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.
Aftermath
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.
Sources
- Cognee raises $7.5M in Seed funding
- Cognee Secures $7.5 Million Seed to Build Permanent Memory for AI Agents
- GitHub 2026-07-13 每月趋势项目
- Show HN: Cognee – Open-Source AI Memory Layer That Remembers Context
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card