Member-only story
Featured
Andrej Karpathy’s Fix for LLM Memory Works on Code Too
Two open-source tools that turn his idea into a code graph for your own codebase.
You ask your coding agent to find every caller of a function before you touch it. It greps the repo, opens four or five files, skims through imports it doesn’t need, and hands you an answer.
Ask again tomorrow, same repo, same question. It starts from zero and does the whole crawl again.
Nothing it learned yesterday carried over.
I wrote about that exact failure in April, just not for code. “RAG, LLM Wiki, or Gbrain? How Your Agent Remembers Changes Everything” compared three ways an agent can remember: search fresh every time, compile a wiki once and keep querying it, or go further and act on what it knows.
Andrej Karpathy’s tweet was the anchor for that piece. His argument: stop treating LLMs as search engines that re-read the same documents on every question. Compile the knowledge into a structure once. Query the structure instead.
That piece was about research and note-taking.
Code has the same problem.
So I went looking for tools that treat a codebase the way Karpathy treats a research…
