Years of memory engineering for AI agents just got beaten by the simplest setup possible 🤯
I read the new paper. Naive RAG beat 12 complex memory systems on 6 out of 7 benchmarks.
The details:
↳ Adding graphs, summaries or atomic facts dropped accuracy by nearly 20 points
↳ Complex systems break the context your agent needs
↳ Naive RAG finished in about 5 minutes, up to 30x faster than EverMemOS
The setup is just raw history, a flat index, and pure evidence handed to the LLM.
We've been making agents think about their memories when plain search was the answer all along.
Why short sessions are doable:
justinmath.com/study-session…
3
A software application or ‘app’ is something that helps you automate some actions consistently that may otherwise require you or human(s) to perform.
7
Kuldeep Singh retweeted
We've been getting a lot of questions lately about how the Internet Archive digitizes books.
The short answer: page by page, by hand.
You may remember our viral 2021 video of Eliza Zhang scanning a book. That's still how we do it.
Meet Eliza, and learn how we scan books: blog.archive.org/2021/02/09/…
If you were to quit your current job and get a new job in the same day paying >$80k, how would you do it?
9
I opened one Google.com tab in Chrome and Safari on my MacBook Air.
Chrome: ~1.3 GB of memory.
Safari: ~281 MB.
Almost 5x difference.
But then I noticed something interesting.
1
1
44
Then there’s the irony.
Safari, which appears to consume significantly less memory, is only available on Macs.
Machines that generally cost more and often come with plenty of memory.
Meanwhile, people using cheaper Windows machines with limited RAM are far more likely to use.
1
1
9








