What Is an LLM Wiki? A Knowledge Base Built for AI Agents

The short definition

An LLM Wiki is a knowledge base that a language model can maintain over time. Instead of only retrieving chunks from raw files when you ask a question, the AI keeps a persistent wiki layer: pages for entities, concepts, summaries, comparisons, open questions, and evolving synthesis.

That distinction matters. In a normal file upload or RAG workflow, the model often rediscovers the same context again and again. It searches, pulls fragments, and assembles an answer for the current prompt. The answer might be useful, but it usually disappears into chat history.

In an LLM Wiki, useful work is filed back into the knowledge base. A new source can update ten existing pages. A good answer can become a new page. A contradiction can be marked where future agents will see it. The system compounds.

The three-layer model

A useful LLM Wiki has three layers.

• Raw sources: the material you collected, such as notes, articles, transcripts, screenshots, PDFs, and project documents.
• The wiki: AI-maintained markdown pages that summarize, connect, and synthesize those sources.
• The schema: the operating manual that tells the AI how to name pages, cite sources, update links, ingest new material, answer questions, and lint the wiki.

The raw sources should stay stable. They are the evidence trail. The wiki can evolve because it is an interpretation layer. The schema is what turns a generic chatbot into a disciplined maintainer.

raw sources -> maintained wiki -> schema-led workflows -> agent memory

This is why the phrase "LLM Wiki" is more precise than "AI notes app". The important part is not that AI can summarize a note. The important part is that AI can maintain a living structure that remains useful after the chat ends.

How it differs from a traditional wiki

Traditional wikis depend on people doing maintenance. Someone must create pages, keep summaries current, add links, merge duplicates, mark stale claims, and decide when a concept deserves its own page. That work is valuable, but it is also the reason most personal wikis decay.

An LLM Wiki changes the maintenance cost. The human still chooses sources, asks questions, and decides what matters. The AI does the repetitive bookkeeping: filing, linking, summarizing, comparing, and checking consistency.

The human role becomes more editorial:

• Which sources should enter the system?
• Which claims deserve trust?
• Which contradictions matter?
• Which questions are worth pursuing next?

The AI role becomes operational:

• Update the relevant pages.
• Add cross-links.
• Keep the index current.
• Record what changed.
• Surface stale or unsupported claims.

That split is the product opportunity. Knowledge management becomes less about manual filing and more about guiding an agent that maintains your context.

How it differs from RAG

RAG is a retrieval pattern. It finds relevant chunks and gives them to a model at query time. That is useful, especially when the source set is large or frequently changing.

But RAG alone does not create a durable synthesis. If five documents together imply a new conclusion, the model may need to rediscover that relationship every time. If a later document contradicts an earlier one, the system may retrieve both without knowing which claim superseded the other.

An LLM Wiki can use retrieval, but it is not limited to retrieval. It keeps a maintained layer where synthesis, contradictions, and links already exist.

RAG:        question -> retrieve chunks -> answer
LLM Wiki:   source -> update wiki -> query wiki -> file useful answers back

The practical result is that your knowledge base becomes easier to ask and easier to inspect. You can read the wiki, follow links, check source references, and see how an answer was built.

What a good ingest looks like

Ingest is the workflow that turns a new source into maintained knowledge. A good ingest does not simply store the file. It updates the wiki.

A single article might create a source summary, update an entity page, add two concept links, revise a comparison page, and append an entry to the log. A single meeting transcript might update project status, decisions, people, risks, and next actions.

At small scale, an index.md file can help both the human and the model navigate. At larger scale, search tools can help. But the key idea stays the same: the source is not just uploaded; it is integrated.

Why Walnut cares about local-first

An LLM Wiki can contain highly private material: unfinished thoughts, personal preferences, drafts, research trails, health notes, project decisions, and the context that makes your judgment yours.

That is why Walnut treats the LLM Wiki pattern as a local-first second brain. The default boundary should start on your device. Online sync, hosted AI, and cloud services should be explicit product choices, not invisible background assumptions.

Walnut's direction is to separate capture, indexing, review, and AI suggestions. Mobile capture can be light. Desktop review can be deliberate. AI can propose structure, but you keep ownership and final judgment.

Where Walnut fits

Walnut is building toward a personal LLM Wiki for AI-era knowledge work:

• Capture fragments before they are lost.
• Preserve source context.
• Extract entities, concepts, and relationships.
• Keep AI suggestions reviewable.
• Turn personal knowledge into a memory layer future agents can call.

The goal is not to replace thinking. The goal is to remove the maintenance burden that keeps personal knowledge systems from compounding.

If you want the product framing, start with the Walnut LLM Wiki page. If you want the memory angle, read about AI agent memory.