Source: gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

#Why This Gist

Karpathy rarely publishes on new architectures without sparking a wave of implementations, and this gist is no exception — dozens of open-source projects have tried to operationalize the idea within weeks. The proposal itself is short but load-bearing: instead of traditional RAG chunk-retrieval, have an LLM incrementally maintain a persistent markdown wiki where synthesis, cross-references, and contradictions are already resolved.

#Key Ideas

#Value for Practitioners

This frames a different end state for agent memory than the prevailing "retrieve more chunks" direction — one worth considering before you invest heavily in vector infra. Read the gist first, then the community debate in the comments: provenance, staleness, and whether "wiki" is even the right term are all unresolved and worth forming your own view on.