We built cognee to give AI agents a better memory.
Today, most AI assistants struggle to recall information beyond simple text snippets, which can lead to incorrect or vague answers. We felt that a more structured memory was needed to truly unlock context-aware intelligence.
We give you 90% accuracy out of the box
The best part is that you can do it in just 5 lines of code!
If you're curious to test it out firsthand, try cognee and give us a ⭐ on GitHub! We’d also love to chat about all things AI memory in our lively Discord – join in to:
Share feedback
Discuss features you want to see next
Learn from our awesome community (+300 members)
If you feel inspired to help shape cognee’s future and build the best AI memory layer out there while sharpening your skills, contribute to our open-source codebase. We have plenty of open issues you can start with!
I love Open-source! Giving AI agents structured memory through knowledge graphs is genius. How do you measure the 90% accuracy claim – is it benchmarked against specific datasets or use cases?
@desmond_ren1 Hi Desmond. We have an entire evaluation framework where we benchmarked cognee using F1, EM, LLM as Judge metrics and Human eval (meaning we went over responses and checked manually) on HotPot. We used standard benchmark, but we also ran benchmarks with our clients.
@vasilije_markovic1 oh that is interesting... for me, mem0 setup was a bit complex (deployment side, tbh mostly due to my stack) so will try cognee and see how it fares for unmess!
Report
This looks exciting, congrats on the launch! Do you have any links/docs to understand the graph generation process and how that fits into LLM workflows? I'd love to gain a deeper understanding of it
I like how there are different SearchTypes, for insights, information, etc…
How do you make the search user-specific? I read on website you personalize to user based on user features but didn’t find it in the docs
I love how cognee bring RAG to the next level! very impressive team and product! Looking forward what is next :) @vasilije_markovic1
Report
I'm impressed by this product, would be a game-changer I think. Just one consideration, what are the trade-offs of 5 lines of codes, how can we achieve such simplicity (Sorry I don't have deep dive knowledge on this).
@tomcao2012 Great question! With 5 lines of code you get our default pipeline that works in most general cases. If you have some specific domain knowledge you need to map, you would need to add your rule set and build your own logic, but we provide tooling for this
cognee
Hey Product Hunt Community! 👋
We built cognee to give AI agents a better memory.
Today, most AI assistants struggle to recall information beyond simple text snippets, which can lead to incorrect or vague answers. We felt that a more structured memory was needed to truly unlock context-aware intelligence.
We give you 90% accuracy out of the box
If you're curious to test it out firsthand, try cognee and give us a ⭐ on GitHub!
We’d also love to chat about all things AI memory in our lively Discord – join in to:
Share feedback
Discuss features you want to see next
Learn from our awesome community (+300 members)
If you feel inspired to help shape cognee’s future and build the best AI memory layer out there while sharpening your skills, contribute to our open-source codebase. We have plenty of open issues you can start with!
Curious how cognee might fit into your business?
📅 Book a 1:1 with me
@vasilije_markovic1 Congrats V. This sounds cool to use and Open Source. Great Combo. Will try for sure!
Manna
I love Open-source! Giving AI agents structured memory through knowledge graphs is genius. How do you measure the 90% accuracy claim – is it benchmarked against specific datasets or use cases?
cognee
@desmond_ren1 Hi Desmond. We have an entire evaluation framework where we benchmarked cognee using F1, EM, LLM as Judge metrics and Human eval (meaning we went over responses and checked manually) on HotPot. We used standard benchmark, but we also ran benchmarks with our clients.
For the standard benchmark, results and how to replicate them check here: https://github.com/topoteretes/cognee/tree/main/evals
unmess
This is amazing. How does this compare against something like Mem0? Hope I'm not mixing things up too much hahaha
cognee
@periodawindsy Hi Anil! Have a look at our evals here, we compared against Mem0 and our tool is more accurate -> https://github.com/topoteretes/cognee/tree/main/evals
unmess
@vasilije_markovic1 oh that is interesting... for me, mem0 setup was a bit complex (deployment side, tbh mostly due to my stack) so will try cognee and see how it fares for unmess!
This looks exciting, congrats on the launch! Do you have any links/docs to understand the graph generation process and how that fits into LLM workflows? I'd love to gain a deeper understanding of it
cognee
@keremakaynak Have a look at https://docs.cognee.ai/
Happy to add more info if you are missing something
Tyce
AIThumbnail.so
I love how cognee bring RAG to the next level! very impressive team and product! Looking forward what is next :) @vasilije_markovic1
I'm impressed by this product, would be a game-changer I think. Just one consideration, what are the trade-offs of 5 lines of codes, how can we achieve such simplicity (Sorry I don't have deep dive knowledge on this).
cognee
@tomcao2012 Great question! With 5 lines of code you get our default pipeline that works in most general cases. If you have some specific domain knowledge you need to map, you would need to add your rule set and build your own logic, but we provide tooling for this
@vasilije_markovic1 ah I see bro, thanks for your response.