llm-wiki

llm-wiki

llm-wiki is a Claude Code plugin (and a small Python CLI) that helps you keep a personal knowledge vault next to your projects. You capture sources into raw/, curate linked markdown in wiki/, and optionally generate a static viewer, wire MCP search, or turn on session memory—so your agent has a durable place to read and write, not a one-off chat dump.

It turns scattered source material into a maintained wiki your agent can keep improving over time—slash-command first (e.g. /llm-wiki:setup, /llm-wiki:ingest, /llm-wiki:build-og), without living in flag hell. Quick setup (docs site): skinnnyjay.github.io/wiki-llm/index.html#setup.

This repository is the plugin: commands, skills, templates, and bin/llm-wiki. After setup, your vault usually lives at ./llm-wiki/ inside whatever repo you chose (terminology is at the top of docs/QUICKSTART.md).

Inspiration

If I have seen further it is by standing on the shoulders of Giants.

— Isaac Newton, letter to Robert Hooke (1675).

The workflow borrows from Andrej Karpathy’s “LLM Wiki” gist and Karpathy on X—a simple pattern for turning sources into maintained notes—and from ideas in the MemPalace line of work (Milla & Ben on X; more context). The goal here is a concrete plugin for Claude Code (and friends) with ingest, validation, and agent-facing skills—not a generic “memory product.”

Feature list (vault, session memory, MCP recall, benchmarks, and more): docs/INSPIRATION.md. Alternate install paths and scripted examples: docs/INSTALL.md.

The wiki

We treat the vault as sources first, then curated notes—not one undifferentiated pile of markdown.

• raw/ — Ingested captures (files, URLs, feeds, APIs) land here with explicit paths; optional raw prepare cleans messy HTML/PDF/OCR before you merge.
• wiki/ — The maintained layer: wikilinks, wiki/index.md, wiki/log.md, and topic pages. wiki-ingest and wiki-maintainer help fold raw/ into structure without losing coherence.
• Ship — Validate, wiki-lint, optional vault git with lifecycle-tagged commits, then build-site for a static viewer (serve over HTTP, not file://).
• See the shape — On-demand wikilink graphs, optional knowledge-graph triples, and MCP search (BM25 by default; optional semantic/hybrid) so agents can navigate what you stored.
• Go wide — Topic research and batch loops (/llm-wiki:research, wiki-research-loop) plan sources → raw/ → wiki/ with logging.

Evidence, trust, and why raw/ and wiki/ differ: ETHOS.md.

• Ingest risk — Scraping or ingesting URLs copies arbitrary text into raw/ and later into model context. That includes prompt-injection patterns and bad-faith pages meant to mislead automations or readers—use at your discretion. This plugin does not sanitize the web for you. See skills/references/access-sources-disclaimer.md and ETHOS.md — Ingestion and security.

Memory

We split vault knowledge, chat continuity, and retrieval evaluation—so “memory” stays understandable and under your control.