Ontology learning heads into the open. We crossed 30,000 GitHub stars this month, and Vasilije wrote down the path from hand-built ontologies to ontologies cognee learns from its own traces. Star on GitHub
Founders note
Agents leave a trail behind them: what they tried, which tool they called, what broke, what worked in the end. Storing that trail is easy. Getting something useful back out of it is not, because none of it was written down to be knowledge in the first place.
What helps is an ontology: a shared vocabulary for the things in your world and how they relate to each other. Give cognee one today and it shapes everything the system takes in. Soon cognee will be able to learn that vocabulary from the traces themselves, and then memory and structure start improving each other. Vasilije wrote that path down this week, all the way to the point where people start saying "world model". He is careful about where the honest claim stops.
We also crossed 30,000 stars on GitHub. Eight months ago we were at 10,000; cognee is now one of fewer than 1,200 repositories above 30,000. Thank you to everyone who ran it and told us what they found. Further down: a recap of the hack night we ran with Qdrant in Berlin.
What's new at cognee
Where does memory end and a world model begin?
Once every run is described in the same vocabulary, runs that looked nothing alike become comparable. Comparing them shows which steps tend to lead where. That is the point at which memory stops only recalling and starts anticipating. Vasilije walks the whole path, from ontologies written by hand to ontologies a system learns for itself.
What it demonstrates: memory and structure that improve each other are not a world model: they are the groundwork you would need before building one.
Milestone
30,000 stars on GitHub
The repo crossed thirty thousand stars this month, eight months after it passed ten thousand. Behind that number sit 252 contributors and 2,965 forks, on an Apache-2.0 licence that is not going to change. Thanks to everyone who ran it and told us what broke.
Past events
Building a company brain in one evening
cognee and Qdrant took over a room at Paul-Lincke-Ufer for an evening, and the build was smaller than it sounds: point cognee at a Slack workspace, let Qdrant carry the vector search, and the archive stops being a search box and starts being a company brain. The difference shows up in the ordinary case: instead of hunting for the thread where someone explained why a customer churned, you ask, and the answer arrives with the messages it came from. Teams demoed in two minutes each, and everyone who came went home with cognee Cloud credits.
Come and meet us
Community hours are an open call where we work through cognee questions and whatever you are building with it. No slides, bring a problem.
- Dave's online community hours. Dave runs these as a regular open slot. Bring whatever you are building with cognee; the next date and the topic go out in Discord beforehand. Save your spot
- GTM Engineering Talks: AI Agents #15 — postponed, new date coming. An AI Alliance SF meetup at Digital Jungle SF that we co-host with Neo4j. Doug Skinner on signals for GTM engineers, a talk from Ashkon Yasseri, then an open mic where everyone gets thirty seconds. Check for the new date
- Sep 3, 18:00–20:30 CEST — SLMs and agent memory, with Fastino Labs. Agent memory quality depends on extraction quality: if extraction is slow, expensive or imprecise, retrieval, reasoning and recall all inherit the problem. An evening in Berlin on that layer, hosted by Fastino Labs and cognee. Registration needs host approval. Save your spot
Know someone using agents?
Forward this issue to a builder who cares about self-improving AI memory.