Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

cookyman74/llmwiki-harness

28PythonMIT

A tool that turns your Obsidian notes folder into a self-maintaining wiki an LLM keeps organized -- and even lets forget things on purpose

llmwiki-harness installs a set of Claude Code agents, skills and scripts into an Obsidian vault so that a person only needs to drop source files into a raw folder while the LLM handles summarizing, cross-linking, flagging contradictions and tidying up. Unlike typical RAG setups that re-search from scratch for every question, it promotes information through four memory tiers (working, episodic, semantic, procedural) so the wiki accumulates value over time. The latest version (v2.1) makes ingestion lightweight and defers heavier merging and judgment work to a periodic 'lint' pass, cutting cost and latency substantially.

What it does

  1. Dropping a document into the raw folder and asking to 'ingest' it makes the LLM create a lightweight evidence summary (L2) and draft-only new concept pages (L3); merging with existing pages, confidence scoring and relationship mapping are deferred to a later batch 'lint' step.
  2. Information is organized into four memory tiers modeled on human memory (session scratch, session summary, facts/concepts, repeated procedures); a claim repeated 3+ times gets promoted to a 'fact' page, and a procedure observed 2+ times gets promoted to a 'procedure' runbook automatically.
  3. Each piece of information gets a confidence score from 0 to 1 based on how many sources back it and what kind (official docs vs. spoken/meeting notes), and unconfirmed information decays exponentially over time like the Ebbinghaus forgetting curve, getting flagged for review rather than deleted.
  4. To avoid re-scanning the whole wiki on every question, the system finds keyword-matched seed pages and expands just one link-hop outward, which measured a roughly 55% reduction in tokens used for fact-briefing type queries compared to the old full-index-read approach.
  5. Routine tasks like ingestion are handled by a cheaper model (sonnet) while judgment-heavy tasks like synthesis and auditing use a more capable model (opus), which the authors report reduced real cost per page by about 4.6x and latency by about 3x.

Why it matters

It tackles the common problem of notes piling up without ever getting organized by having an LLM act as the ongoing maintainer, and it offers a concrete, reusable open-source framework that can be applied directly to personal knowledge management, meeting note processing, or team wiki operations.

Terms in this repo

  • RAG · A method that searches relevant document snippets for every question rather than accumulating reusable knowledge over time
  • Frontmatter · A metadata block (title, tags, confidence, etc.) placed at the top of a markdown file
  • Ebbinghaus forgetting curve · A theory that memory retention decays exponentially over time, applied here to calculate how 'fresh' stored information is
  • Confidence score · A 0 to 1 rating for a claim based on how many sources support it and what type of sources they are
  • Supersession · Marking outdated information as 'stale' and linking it to a replacement page instead of deleting it, preserving a history of changes

Repository description (English)

cookyman74/llmwiki-harness

Open on GitHub

Trending repos

All repos →

Latest from METAL MEDIA