🧠

Knowledge Wiki

A persistent knowledge base that compounds over time. Capture sources, maintain topic pages, and track open questions — your AI does the bookkeeping so cross-references, contradictions, and synthesis stay up to date without you lifting a finger.

Requires an AI connected to your Hjarni account via MCP.

Structure

Tags

source topic question resolved

Folders

Knowledge Wiki
A living wiki maintained by you and your AI
Has AI instructions
3 starter notes
Sources
Original material and AI summaries — the raw input is always preserved
Has AI instructions
Topics
Maintained wiki pages — one note per subject
Has AI instructions
Open Questions
Things to investigate or think about
Has AI instructions
Changelog
Log of what the AI updated and why
Has AI instructions

For your AI

Share this page with Claude. It reads the definition below, creates the folders, instructions, tags, and starter notes in your account.

Install steps for AI agents:

  1. Check existing tags with tags-list. Only create missing ones with tags-create.
  2. Create containers top-down using containers-create, noting the returned IDs. Use parent_id to build the hierarchy.
  3. For each container with llm_instructions, call instructions-update with level: "container" and the container's ID.
  4. Create any seed notes using notes-create, placing them in the correct container by ID. Use container_path to resolve which container.
  5. Discuss any customizations with the user before or after installing.
---
name: Knowledge Wiki
description: 'A persistent knowledge base that compounds over time. Capture sources,
  maintain topic pages, and track open questions — your AI does the bookkeeping so
  cross-references, contradictions, and synthesis stay up to date without you lifting
  a finger.

  '
tags:
- source
- topic
- question
- resolved
containers:
- name: Knowledge Wiki
  description: A living wiki maintained by you and your AI
  llm_instructions: |
    This container is a persistent knowledge wiki. The goal is to build knowledge that compounds — not to re-derive answers from scratch every time.
    - Before creating any note, search the entire Knowledge Wiki for existing notes on the same subject. Update existing notes instead of creating duplicates.
    - Read the "Wiki Guidelines" note in this container before doing any work. It defines how this wiki is maintained.

    Ingest workflow (when the user shares a new source):
    1. Create a note in Sources with the original material and a structured summary. Save the URL in source_url when one exists.
    2. Check every existing Topic note that the source touches. Update them in place with the new information.
    3. If the source covers a subject with no Topic yet, suggest creating one.
    4. Update the Index note with any new Topics or Sources created.
    5. Add an entry to the Changelog for each Topic that was updated.
    A single source may touch many Topic notes. That is expected.
    If the user is actively in conversation, ask what stood out or what to emphasize. If they dropped a link without comment, complete the full ingest without blocking on a question.

    Query workflow (when the user asks a question):
    - Read the Index note first to find relevant Topics, then drill into them. Fall back to search if the Index does not cover the question.
    - Synthesize an answer from existing wiki pages with links.
    - If the answer required real work (comparison, analysis, synthesis across sources), suggest filing it as a new Topic so it compounds in the wiki instead of disappearing into chat history.

    Lint (periodic health check — do this when the user asks, or suggest it occasionally):
    - Look for: stale claims superseded by newer sources, Topic notes with no linked Sources, Sources that were never synthesized into Topics, Open Questions that existing Topics already answer, and missing cross-references between related Topics.
    - Report findings to the user and suggest fixes. Do not bulk-edit without approval.

    General rules:
    - When updating a Topic note with new information, add a short entry to the Changelog so the user can see what changed and why.
    - When a Topic note contradicts new information, do not silently overwrite. Flag the contradiction in the Topic note and mention it to the user.
    - Link related notes using wiki-style references. Every Topic links to its Sources. Related Topics link to each other.
    - Keep note titles short and specific. Use noun phrases, not sentences. "SQLite Write-Ahead Logging" not "How SQLite WAL Mode Works and Why You Should Use It".
  children:
  - name: Sources
    description: Original material and AI summaries — the raw input is always preserved
    llm_instructions: |
      Use this folder to capture inputs before they are synthesized into Topics.
      - Create a note here when the user shares an article, link, quote, conversation, or idea.
      - Tag every note with "source".
      - Save the original URL in source_url when one exists.
      - Structure every source note in two sections:
        **Original** — the raw material, preserved verbatim. For a URL source this can be a link; for a quote, paste, or conversation excerpt, copy the full text here unchanged. This section is immutable — never edit it after creation.
        **Summary** — an AI-written structured summary below: key claims, main arguments, data points, and why it matters. Be faithful to the original — do not editorialize or interpret.
      - Do not edit source notes after initial creation unless correcting a factual error against the Original section. New interpretations and synthesis belong in Topic notes, not here.
      - If in doubt about a claim during later work, re-read the Original section rather than trusting the Summary alone.
      - After creating a source note, follow the ingest workflow in the parent container instructions: update existing Topics, suggest new ones, log changes.
  - name: Topics
    description: Maintained wiki pages — one note per subject
    llm_instructions: |
      Use this folder for synthesized, maintained knowledge pages.
      - Each note covers one subject. Keep the scope narrow enough that the note stays useful.
      - Tag every note with "topic".
      - Structure topic notes with: a one-paragraph summary at the top, then sections for detail. End with a "Sources" section linking to the source notes that informed this page.
      - When new information arrives, update the relevant section in place. Do not append to the bottom in a log format — this is a wiki page, not a journal.
      - If a topic grows too large, split it into sub-topics and link them together. Ask the user before splitting.
      - When two Topic notes overlap significantly, suggest merging them.
      - Flag contradictions clearly: "Note: [source X] says A, but [source Y] says B. This is unresolved."
  - name: Open Questions
    description: Things to investigate or think about
    llm_instructions: |
      Use this folder for questions the user wants to explore.
      - Create a note here when the user raises a question that cannot be answered from existing Topics.
      - Tag every note with "question".
      - Keep question notes short: the question, why it matters, and any leads or partial answers.
      - When a question is answered (by a new source or the user's own thinking), move the answer into the relevant Topic note. Then either delete the question note or move it to Topics — tag it "resolved" and remove the "question" tag.
      - If a new source partially answers an open question, update the question note with the partial answer and link to the source.
  - name: Changelog
    description: Log of what the AI updated and why
    llm_instructions: |
      Use this folder to maintain a running log of wiki updates.
      - Create one note per month, titled with the month and year (e.g., "April 2026").
      - Each entry is one line: the date, what changed, and why. Keep entries to one or two sentences.
      - Log every Topic update that was triggered by a new Source. Do not log minor formatting changes.
      - This log helps the user see how their wiki is evolving without having to diff every note.
  notes:
  - title: Index
    body: |
      A catalog of everything in this wiki. The AI reads this first when answering questions or running lint, then drills into relevant notes. Updated on every ingest.

      ## Topics

      *(none yet — topics appear here as you add sources)*

      ## Sources

      - [[Example Source]] — Karpathy's LLM Wiki pattern: persistent wikis over RAG

      ## Open Questions

      *(none yet)*
    tags:
    - topic
    container_path: Knowledge Wiki
  - title: Wiki Guidelines
    body: |
      This note defines how your Knowledge Wiki works. Your AI reads this before doing anything.

      **Core principle:** This wiki compounds. Every new source should make existing pages better, not just add another note to the pile.

      **Rules for the AI:**
      - Search before creating. If a topic exists, update it. Do not create duplicates.
      - One note per topic. Keep it focused. Split when it gets too long.
      - Link everything. Every Topic links to its Sources. Related Topics link to each other.
      - Flag contradictions. Do not silently overwrite — surface the conflict and let the human decide.
      - Log changes. When a Topic is updated because of a new Source, add a line to the Changelog.
      - Capture first, synthesize second. New inputs go to Sources before they become part of a Topic.
      - File good answers back. When a question produces a useful analysis, save it as a Topic so it compounds.
      - Lint occasionally. Look for orphan notes, stale claims, and missing cross-references.

      **Rules for you:**
      - Drop anything here — links, quotes, half-formed ideas. The AI will file it.
      - Review the Changelog occasionally to see how your knowledge is growing.
      - Edit Topic notes directly whenever you want. This is your wiki, not the AI's.

      Edit these guidelines to match how you think and work. These are starting defaults.
    tags:
    - topic
    container_path: Knowledge Wiki
  - title: Example Source
    body: |
      ## Original

      *URL source — see source_url for the full gist.*

      ## Summary

      **Type:** Article
      **Author:** Andrej Karpathy
      **Published:** April 2026

      Proposes a pattern called "LLM Wiki": instead of RAG (re-deriving knowledge on every query), have an LLM incrementally build and maintain a persistent wiki of interlinked markdown files.

      **Key claims:**
      - RAG rediscovers knowledge from scratch every time. A maintained wiki compounds — cross-references are already there, contradictions already flagged, synthesis already done.
      - The bottleneck in knowledge management is not reading or thinking — it is bookkeeping. Humans abandon wikis because maintenance grows faster than value. LLMs do not get bored.
      - Three layers: raw sources (immutable), the wiki (LLM-maintained markdown), and a schema document that defines conventions and workflows.
      - Three operations: ingest (process a source into the wiki), query (answer questions from the wiki), lint (periodic health check for stale claims, orphans, and gaps).

      **Connections:** Vannevar Bush's Memex (1945) — a personal knowledge store with associative trails. The missing piece was who does the maintenance.

      *This is an example note — feel free to delete it.*
    tags:
    - source
    source_url: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
    container_path: Knowledge Wiki > Sources

Write once. You both remember.

Free to start. No credit card required.

Start writing free

Works with Claude and ChatGPT today. Gemini coming soon.