MementoALTERNATIVES / LLM WIKI
Karpathy’s LLM wiki, without running it yourself
In April 2026 Andrej Karpathy published “LLM Wiki”, an idea file: keep your raw sources in a folder, and let a coding agent write and maintain a wiki of Markdown pages from them. Memento is that idea as a hosted product. Here is what running it yourself does better, what Memento adds, and when to build your own.
CHECKED AGAINST KARPATHY’S ORIGINAL WRITE-UP AND THE PROJECTS’ READMES ON 15 SEPTEMBER 2026
The pattern, in Karpathy’s words
The gist describes three layers. The raw sources are “immutable — the LLM reads from them but never modifies them.” The wiki is a directory of Markdown pages, and “the LLM owns this layer entirely.” A schema file — CLAUDE.md for Claude Code, AGENTS.md for Codex — tells the agent how the wiki is organised. You then run three operations: ingest a source (“a single source might touch 10–15 wiki pages”), query the wiki, and lint it (“periodically, ask the LLM to health-check the wiki”).
The case for it is one sentence: “Humans abandon wikis because the maintenance burden grows faster than the value.” An LLM does not get bored of the maintenance, so the wiki becomes, in his words, “a persistent, compounding artifact.” And in the tweet that started it, Karpathy added: “I think there is room here for an incredible new product instead of a hacky collection of scripts.” Memento is our attempt at that product.
What running it yourself does better
It is entirely yours. The gist is “intentionally abstract” — the idea, not an implementation — so you choose the agent (Karpathy names Codex, Claude Code and OpenCode), the model, the directory layout and every rule in the schema. The files stay on your machine, and you watch the agent work in real time.
Obsidian is the viewer. “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.” You get the graph view, the Web Clipper for grabbing articles, and plugins like Dataview on top of the pages the agent writes.
No limits but your own budget. There is no daily run allowance and no queue: a change is on disk the moment the agent writes it, and you can ingest one source at a time and stay involved, which is how Karpathy says he prefers to work.
You don’t have to start from nothing. Open-source implementations have appeared since — among them nashsu/llm_wiki (a desktop app, GPL-3.0), claude-obsidian (slash commands for Claude Code, MIT) and llm-wiki-agent (Claude Code, Codex, OpenCode or Gemini CLI, MIT).
What Memento does differently
It runs without you
In the gist, all three operations start when you ask for them — you ingest, you query, you periodically ask for a lint — and the popular implementations run as commands in a local agent session. Memento’s agent is hosted next to your library. You share an article from your phone, and it is filed by the time you sit down; a routine runs a health check or a weekly sweep on a schedule, with nothing of yours switched on.
Humans in the loop, built in
For a business wiki, Karpathy suggests “possibly with humans in the loop reviewing updates”. Memento makes that the default for anyone. Changes that only add — filing a capture, linking pages, appending a sourced claim — go straight in. Changes that could lose or distort what you already trust — rewriting a confirmed fact, merging pages, resolving a conflict between sources — wait in your inbox as a diff until you approve them. See the full list.
Every fact has a status and a receipt
The gist’s lint step flags contradictions. Memento labels every fact when it is written: confirmed, inferred, disputed or outdated. An inferred fact is never silently promoted to confirmed, and each claim cites the read-only source page it came from — Karpathy’s immutable raw layer, linked claim by claim.
Nothing to operate
No agent to install, no schema to write before the first page appears, no machine to keep on. Memento ships the page format and the filing rules, and you add house rules per space under Instructions. The agent runs in a private sandbox with no network access until you approve a destination, on your own Claude, OpenAI or other model account.
The same files underneath
Karpathy: “The wiki is just a git repo of markdown files.” So is Memento’s. Every change, yours or the agent’s, is a commit; the export is plain Markdown with wikilinks, your original files and the full history. You can open it in Obsidian and carry on with the DIY version at any time.
What each costs
Running it yourself costs your model usage and your time. Memento costs the same model usage — it runs on your own account — plus a plan: free for one knowledge space of your own and 20 agent runs a day, or €10 a month for 25 spaces and 200 runs a day (see Plans & limits). Asking questions never counts against either.
Build your own instead if …
… you want the files never to leave your machine, you enjoy designing the schema and tuning the agent, you already live in a terminal with a coding agent open, or you want to experiment with the pattern itself. Memento is for the person who wants the result — a wiki an agent keeps current — without operating it. Filing a capture takes minutes of agent work (typically around eight), not the seconds of watching your own agent write.
Try it on your own notes
Memento is in private beta and lets people in a handful at a time. The free plan is one knowledge space of your own plus the sample space, 20 agent runs a day, and every feature. You bring your own Claude, OpenAI or other model account, so model usage is billed by your provider, not by us.