cognee 1.0 is live. We shipped a major open-source release and a new cloud that allows your agents to learn from their mistakes and self-improve over time. Star on GitHub or read the launch notes.
Welcome
We're glad to welcome you to the first cognee newsletter. Our plan is simple: one memo every two weeks. You'll learn about new product features, frontier research and POCs of our AI lab, and the daily struggles of building a startup between SF and Berlin. This is also where we share news about our upcoming community events and hackathons across the globe.
Ultimately, our goal is to build a community of like-minded people, excited about AI memory, sharing knowledge and supporting each other building agents.
Watch: cognee 1.0 launch video
If you want the quick version, start here. It's the cleanest overview of what changed in 1.0 and why we think memory needs a better abstraction than just more context.
What's new
Some honesty about how we got here: First, we started building memory on top of a graph database. That was powerful, but too hard to manage for an individual developer. Second, no serious team would hand memory to a black box which they couldn't inspect, nor take with them if they left. Those two lessons are why cognee 1.0 looks the way it does.
- cognee 1.0: the new API, in four verbs. Instead of pipeline steps, cognee now thinks the way an agent does: remember, recall, improve, forget. The genuinely new one is improve: memory re-weights itself based on how agents actually use it. How it works under the hood.
- One Postgres instance is enough. The entire memory layer, graph, vectors, all of it, now runs self-hosted on a single Postgres instance. No new infrastructure and no graph database to operate.
- cognee Cloud: memory without operations. If you don't want to touch Postgres: Cloud is the fully managed version. Create an account, grab an API key, done. You can pair Cloud with a local setup for persistent, always-on workloads.
- The new Rust SDK: memory on the device. With the Rust core, cognee now runs on-device and at the edge — memory where your agents are, without a round trip to the cloud.
- BEAM: the numbers behind it. On the BEAM benchmark, cognee scores 79% at 100K context and 67% at 10M context. The best previously reported results were 73.4% and 64.1%. Token usage stays flat as your data grows.
You can read more in the 1.0 announcement.
Launch week: AI Engineer San Francisco
We did our launch in June and walked into AI Engineer World's Fair a few days later. Not the calmest week we've had, but probably one of the most energizing. Here's a quick look at launch week and the meetup around it.
Warsaw hackathon: build a GTM brain with cognee
Our CEO, Vas, led an interactive session with hundreds of builders, showing them how to build a GTM brain with cognee.
Connect HubSpot, Granola, Slack, email, and the rest of your sales intelligence into one unified memory layer that your agents can use. No more fragmented context across sales reps, missing information, or missed follow-ups.
Know someone building agents?
Send them this memo. Good memory infrastructure gets better when more builders stress-test it, question it, and use it in the wild.