What LLM Wiki is about
Andrej Karpathy recently described a really nice methodology for knowledge work, called LLM Wiki: instead of dumping everything into one growing chat, keep a personal wiki on disk. Many small markdown pages, each focused on one entity or topic, cross-linked with [[wikilinks]], with an index.md as a directory. The LLM does the bookkeeping; you curate.
Karpathy puts the heart of the pattern very precisely:
The tedious part of maintaining a knowledge base is not the reading or the thinking. It’s the bookkeeping.
That bookkeeping is what quietly retires most personal wikis after a few months. Cross-references pile up, versions drift apart, contradictions scatter through the pages, and at some point nobody comes back to clean it up. LLMs don’t get tired of that kind of chore. Hand them the cross-references, version tracking, contradiction detection, and synthesis, and your knowledge can settle into a form that both you and an AI can keep re-reading and re-shaping over the long run. That’s a fundamental move in the LLM Wiki methodology.
What Nowledge Mem has been doing this past year
Before Karpathy’s writeup went out, Mem had already been working on something close to it. We arrived from a slightly different angle: the entry point on our side was AI tooling. Two long-standing pain points kept compounding: sessions that forget everything once they end, and the very real difficulty of getting any single person to keep a long-running knowledge base alive by hand. Together those two were what kept AI tools from accumulating any real knowledge from their own use. Karpathy’s writeup gave the problem and its core solution a clear method and a clear name.
Over this past year, Mem has been doing exactly that: capturing the knowledge that comes out of your daily AI work and pulling it into one place. Memories you save, sources you import, conversations the browser extension brings back, decisions the agents distil, all stored as connections on one super knowledge graph. Background Intelligence detects topic clusters, evolution chains, and contradictions. Search runs across all of it.
What Karpathy actually points at is the methodology of organising these as a wiki and sitting down to read them as one. We try to take care of the bookkeeping that can be automated; the parts that genuinely need a person, like picking sources, asking the right questions, and deciding what’s worth keeping, stay with you.
What we used to lack
If you read the LLM Wiki post and then opened Mem, the gap was easy to feel. Knowledge, raw sources, the distilled conclusions, and all the connections between them were already in the knowledge graph, but the reading surface wasn’t wiki-shaped at all. You could explore in the Knowledge Graph Agent or the Graph View on your own, but you couldn’t read it the way Wiki Link Surfing lets you, just clicking your way through ~
v0.8: full embrace of LLM Wiki
In v0.8, we redesigned this end of the product so both human users and Agent users get the full-form LLM Wiki.
What we’ve added on the wiki side is built on top of the communities, entities, and crystals that already live in the knowledge graph: the Library now carries a full reading structure shaped like a wiki.
What used to be a Library of documents and URLs is now a wiki you can actually read through:
- The Wiki tab sits as the leftmost tab on the Library. It groups your knowledge by topic clusters the AI has already named for you. Each card lists the concepts most-discussed inside that cluster, plus a few of the durable Crystals that summarise them. Click any card and the right pane opens its topic page.
- An entity page is that entity’s wiki entry: which topic it belongs to, when it was first identified, the AI’s confidence score, any aliases it picked up, the memories that mention it, and the entities sitting around it. All of this is derived live from the graph; you don’t have to maintain a separate copy.
[[Entity name]]wikilinks are clickable now, both in Crystals and in AI-generated artifacts. The agent writes them on its own too, so newly distilled content arrives pre-linked. Case-insensitive, alias-aware.- “Built from N memories” sits at the bottom of every Crystal page and AI-generated artifact, listing the source memories ranked by their contribution. Click any row to peek the underlying memory in place.
- Study with AI turns any source into a guided extraction: the AI reads the whole source, cross-checks against memories you’ve already saved, and surfaces a few worth keeping (memories or Crystals, with wikilinks already written in). Underneath there are three buttons: Save 1, Save all, Skip. No need to type a reply.
- Wiki Export ships your whole wiki as a portable folder of plain markdown (
index.md,topics/,entities/,crystals/). The same wikilinks resolve in Obsidian, Logseq, or any markdown reader. Re-export whenever you want a fresh copy. - The single-page wiki API lets agents and scripts grab one page (entity, crystal, or topic) as markdown over HTTP, without pulling the whole wiki.
- The right column on Crystal and Topic pages got real. Crystal pages now show an Evolution section: which older memories grew into this one, and which newer ones it grew into. Click any row to read the original. Topic pages carry an “Also relates to” rail so you can hop to a neighbouring cluster without going back to the index. Every wiki page now has a “Last touched” date; if it would be the same day as “First seen”, we just show one.
- Topic pages also show what’s recent on the topic. A “Recent on this topic” rail lists the latest memories that touch any of its entities, answering “what have I been thinking about on this topic lately” much faster than flipping through the whole wiki.
- Wiki pages now show the graph. Entity, Crystal, and Topic pages each embed a small graph on the right. It’s the same graph the Graph tab renders, so nodes are clickable, zoom and hover work as expected. The Topic version is the interesting one: it shows the cluster’s actual shape, which of its entities link to which, plus the memories that sit at the centre because they touch the most of them. Even on a 200-entity topic, what you see is the cluster’s real outline.
- Investigate, from a wiki page into the Knowledge Graph and Graph Intelligence. Every wiki page has an Investigate button now. On an entity or crystal page it opens the Knowledge Graph with that node already selected. On a topic page it highlights the whole cluster at once. From there you can pan, expand a node, or hand the cluster to Graph Intelligence (Mem’s reasoning agent) and ask it what’s going on, where it contradicts itself, or to draft a brief. Anything Graph Intelligence saves comes back to the same topic page in Library.
Nowledge Mem’s LLM Wiki
Nowledge Mem’s LLM Wiki layers the “reading path” Karpathy described on top of the “writing path” Mem already had. Background Intelligence keeps doing the same work it always did on the super knowledge graph (conflict detection, evolution, bi-temporal timelines, forgetting). The new Wiki finally gives you a place to actually read what comes out of it.
A few boundaries we’re keeping on purpose:
- Wiki pages are projected live from the data layer (memories, sources, crystals). Edits happen at the data layer, and pages refresh from there.
- Export gives you a one-shot snapshot copy. Re-export whenever you want a fresh one.
- The AI only proposes. You decide what to keep. “Study with AI” never auto-saves.
Try it
- Open the Library and switch to the Wiki tab. If it’s empty, open the Graph view, run community detection once, and come back.
- Pick a topic that already has a few memories. Read the cluster, click into an entity, then into a related entity. You should be able to navigate the whole way without typing.
- Open any source and click “Study with AI”. Read the proposals, tap Save 1 on something useful. Find it in the Memories view.
- From any wiki page, click Investigate. The Knowledge Graph opens with the entity, crystal, or whole topic already selected. From there you can pan, expand neighbours, or hand the cluster to Graph Intelligence and ask it what’s actually going on, what contradicts what, or to draft a first pass. Anything you keep comes back as a Crystal in the same Wiki page you just left.
- Click Download on the tab row. Open the ZIP in Obsidian or any markdown reader. The same wikilinks should still render and resolve there.
Credit
Karpathy’s LLM Wiki has been a useful mirror for Nowledge Mem. It showed us what we’d been missing on the “knowledge being read” side.
What’s next
v0.8 is the first release where the reading side of Mem actually catches up with the writing and self-evolving side.
From here, we want to tighten the agent loop further inside the Wiki itself, for example:
- have the conversations Investigate leaves behind seed the originating cluster as a Crystal starting point in Library, automatically,
- give external sources like Apple Notes and Obsidian a two-way channel so changes you make outside Mem flow back in, without manually re-importing,
- have the EVOLVES timeline weave together the different versions of the same concept across time and sources into one readable trajectory.
Download 0.8.0 · Library docs · Read the original LLM Wiki writeup