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
- Three layers: immutable raw sources, an LLM-maintained wiki of interlinked markdown pages, and a schema file (CLAUDE.md / AGENTS.md) governing LLM behavior
- Three operations: ingest (process new sources), query (answer and file good answers back into the wiki), lint (periodic health checks for contradictions and gaps)
- The wiki is a compounding artifact, not a retrieval cache
#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.