Skip to content

What is the llm wiki Template

Jump to bottom
Chris Sweet edited this page Jul 14, 2026 · 1 revision

type: index up: "Home_llm-wiki-KB" related:

What is the llm-wiki Template

The llm-wiki-memory-template is a project template that gives a long-running project a small, LLM-maintained, cross-linked Markdown wiki as durable memory, so an agent's understanding of the project compounds across sessions instead of being thrown away at the end of each conversation. This page is the orientation anchor for the rest of this bundle.

What is the llm-wiki-memory-template?

It is a GitHub template repository at crcresearch/llm-wiki-memory-template. You instantiate a new project from it, and the project gains a persistent wiki plus the conventions, agent overlays, and tooling that keep that wiki maintained. The template packages the llm-wiki pattern (Karpathy lineage) and adds a discipline layer on top: agent overlays, a verification gate, a typed-edge ontology, attribution-aware logging, and a knowledge-graph pipeline. See Template-Philosophy for why it is built this way.

What is an llm-wiki?

An llm-wiki is a directory of interlinked Markdown files with YAML frontmatter that an LLM writes and maintains, and a human reads and directs. It sits between the raw source material and the human, holding the compiled, cross-referenced understanding of a project. The human curates sources, explores, and asks questions; the LLM does the summarizing, cross-referencing, filing, and bookkeeping. The wiki is just a git repo of Markdown, so it comes with version history, branching, and collaboration for free.

What problem does the template solve?

Ordinary LLM-over-documents workflows (RAG, NotebookLM, file uploads) rediscover knowledge from scratch on every question. Nothing accumulates. Ask a subtle question that requires synthesizing five documents, and the model has to find and piece together the fragments every time. The template's answer is a persistent, compounding artifact: the LLM reads a source once, integrates it into the wiki, and the cross-references, contradictions, and syntheses are already there for every future question. The tedious part of a knowledge base is not the reading, it is the bookkeeping, and LLMs do not get bored of bookkeeping. See Template-Philosophy.

What does the template give me when I install it?

Instantiation scaffolds a working project. Concretely you get:

  • A wiki sub-repo at wiki/<repo>.wiki/, seeded with Home, index, log, a SCHEMA conventions reference, and the typed-edge vocabulary in Edge-Types.md. See Wiki-Structure-and-Conventions.
  • A generated CLAUDE.md (or AGENTS.md) carrying the memory-boundary and wiki-maintenance rules that make the agent a disciplined wiki maintainer.
  • An agent overlay for your coding assistant. The Claude Code overlay ships slash commands (/wiki-source, /wiki-experiment, /wiki-lint), model-side skills, a SessionStart hook that auto-clones and refreshes the wiki, and an optional advisory PostToolUse hook.
  • Agent-agnostic discipline files under wiki/agents/ (verification gate, discipline gates, write protocol).
  • A knowledge-graph pipeline under scripts/kg/ that turns the wiki's frontmatter and links into a SPARQL-queryable RDF graph. See Knowledge-Graph.

For the step-by-step, see Installation-and-Setup. For what you do afterward, see Everyday-Operations.

How is this different from a normal wiki?

A normal wiki is maintained by humans, and humans abandon wikis because the maintenance burden grows faster than the value: cross-references rot, summaries go stale, contradictions accumulate. In an llm-wiki the LLM does the maintenance, so the cost of keeping it current is near zero. The wiki stays healthy because the boring work is automated, not because someone is disciplined enough to keep doing it.

How is this different from Obsidian, Notion, Roam, Confluence?

Those are human-authoring surfaces: tools for a person to write and organize notes. The llm-wiki inverts the roles. The LLM is the author; the human reads and directs. The template is agent-first, not human-first. It is complementary to Obsidian in particular: a common setup keeps the LLM agent on one side making edits and Obsidian open on the other for browsing the graph in real time. Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase. The template does not replace those tools; it changes who does the writing.

How is this different from NotebookLM or vanilla RAG?

RAG and NotebookLM index raw sources and retrieve chunks at query time. The synthesis happens fresh on every question and then disappears into chat history. The llm-wiki adds a layer where an insight can settle: once filed, a finding is referenced by name rather than rediscovered, and every new page that links to it makes it easier to surface next time. RAG re-derives; the wiki accumulates. The template is not a rejection of RAG (dense retrieval is still the right tool for single-hop lookups over modest corpora), it is a different tool for the case where knowledge should compound. See When-to-Use-When-Not.

Is this for research groups, individuals, or both?

Both. The positioning materials (the WGRA flyer) frame it around research groups, where it acts as durable institutional memory: the "long-serving lab technician" whose knowledge does not walk out the door when a student graduates. But the pattern applies equally to an individual tracking a deep-dive over weeks, reading a book, or running a solo research program. For team use, see Team-Workflow; for the decision of whether it fits your project at all, see When-to-Use-When-Not.

What does "LLM-authored wiki" mean, concretely?

It means the human almost never writes wiki pages by hand. The workflow is: you drop a source into the project or run an experiment and tell the LLM to file it. The LLM reads it, discusses the takeaways with you, writes a summary page, updates the index, revises related pages, fixes cross-references in both directions, and appends a log entry, touching 5 to 15 pages in one pass. You review the result and guide emphasis. The three named operations (Ingest, Query, Lint) are how the LLM interacts with the wiki; see Everyday-Operations and Template-Philosophy.

What does "durable memory" mean here?

Durable memory is knowledge that survives across sessions, users, and machines rather than living only in a single conversation's context window. The wiki is that memory: the agent reads it to recall project context at the start of work and writes to it to remember what a future session would need. Because it is a git repo, the memory is versioned, auditable, and portable. The template's own guidance states the rule directly: read the wiki to recall, write to it to remember, proactively and in both directions. This is distinct from the per-user Claude-memory layer, which holds user identity and preferences; the wiki holds project knowledge (see the Memory boundary in the generated CLAUDE.md).

See also

Clone this wiki locally