Make your AI agents remember better with a Karpathy wiki

I have been using personal AI agents from Meta and X for the last couple of months, and yes, they do have memory. They remember things across chats. But I could not read that memory, check it, or keep it the same way in both agents.

They still flagged tasks as pending weeks after we closed them. They still asked me for facts I had already confirmed. Built-in memory is a summary I cannot see inside. I could not see what it kept, what it dropped, or where it came from. And it did not move from one agent to the other in the same form.

A similar memory issue had been fixed before. The question was whether personal agents like Muse and Grok could use skills, and whether a skill could fix their memory the same way. That is what I was testing.

Wiki to the rescue

I use Astro-Han/karpathy-llm-wiki, a Karpathy-style wiki skill from Andrej Karpathy's llm-wiki pattern. I tested it on both agents, and it works on both. I run the same memory system on Muse, Meta's personal AI assistant, and on my Grok agent.

Their memory stays. It just gets a shape I can read and check: a wiki on top of it. Original sources go in a raw folder and never change. The agent turns those sources into linked wiki pages, keeps an index, and writes every change in a log. It does not write a number, date, or quote unless it finds those exact words in the original source. If a source adds nothing new, it can say so and stop. If a new source contradicts an old claim, it keeps the old claim and marks it Disputed or Outdated, instead of quietly rewriting history.

Do it yourself, the same way.

First, get the skill onto your agent. Tell it to install Astro-Han/karpathy-llm-wiki, the karpathy wiki skill on GitHub, and follow it. That is how Grok did it: it fetched a version from GitHub on its own. If your agent cannot find or install it, give it the skill file directly: paste the repo's SKILL.md into chat and tell your agent this file is the skill, save it and follow it exactly. Do not let it invent its own wiki structure from scratch.

Then paste this to your agent, once. It is a standing instruction, not a one-time task:

Use typed pages: summaries, entities, concepts, and comparisons. Keep one wiki for one person.

Before writing any new page, surface 3 to 5 key takeaways, the people involved, and the existing pages likely affected, and wait for my correction if I give one.

Before you tell me something is pending, missing, or unresolved, check the wiki and our chat history first. Never re-ask me for a fact I already confirmed.

Keep raw sources immutable. Do not write a number, date, or quote unless you can find it in the original source. When a source contradicts an old claim, keep the old claim and mark it Disputed or Outdated, never silently rewrite it. Prefer updating an existing page over creating a near-duplicate.

Do not file a query answer unless I ask you to. After every change you do file, update the index and append to the log, then confirm back with your first log entry.

Then just talk to it. To file something, say: "Ingest this article: [link]". To ask what it knows, say: "What do I know about [topic]?" To check the wiki is healthy, say: "Lint my wiki".

The wiki did not replace the memory my agents already had. It gave that memory a form I can read, check against its sources, and keep the same on Muse and on Grok. I set the rules, the agents do the filing, and I correct them when a takeaway or a page needs correction.