Pinned Tweet
made a vector database that lives inside your app as one file. no server, no index to build.
exact search: 18.8x faster than @lancedb. ready to query: 124x faster than faiss hnsw. filtered search: ~180x faster than chroma.
github.com/a3ro-dev/qenlo
okay this post is going viral.
i tried @theo's method of looping over a problem w/ an agent until it just somehow finds out how to fix it. this wastes tokens. so yeah. but man this works. this really works. i used this inside of claude-code (@t3dotcodes doesn't work for me lol) and it worked. i cannot believe
i tried @theo's method of looping over a problem w/ an agent until it just somehow finds out how to fix it. this wastes tokens. so yeah. but man this works. this really works. i used this inside of claude-code (@t3dotcodes doesn't work for me lol) and it worked. i cannot believe
it isn’t entirely in the eye of the beholder,
some changes are pretty clearly regression on almost every axis, like sliding into addiction, cruelty, or self-destruction, even if the person frames them as freedom.
---
don't take my word for the qenlo numbers, rerun them.
the harness and the test data are in the repo, and the raw benchmark results are published too. 18.8x vs lancedb, 124x vs faiss hnsw, ~180x vs chroma. clone it and check.
github.com/a3ro-dev/qenlo
the agent writing a memory is already an llm. so why does every memory layer run a second one to clean up after it?
what this isn't: @mem0ai , zep and @supermemory report 67-97% on these benchmarks, but that's answer accuracy graded by an llm, with an llm in the loop and a server behind it. that's a different number. mine is retrieval recall on a laptop, no llm anywhere.
ponytail.dev/ what a slop website but what a great agent 🔥
