zosmaai/pi-llm-wiki

What it solves

One-shot RAG (Retrieval-Augmented Generation) often results in ephemeral synthesis where the model searches raw documents but fails to retain insights over time. pi-llm-wiki creates a durable, interlinked knowledge base that compounds as you capture sources and perform analyses, transforming raw data into a structured wiki that persists across sessions.

How it works

The system uses a layered vault architecture (Personal and Project levels) to store knowledge in an Obsidian-compatible format. It processes information through a four-layer pipeline:

1. Raw Source Packets: Immutable captures of URLs, PDFs, and other files.
2. Source Pages: Summaries of what each specific source contains.
3. Canonical Wiki Pages: High-level tracking of concepts, entities, and syntheses.
4. Generated Metadata: Searchable registries and backlinks that keep the wiki navigable.

It integrates as an extension for the pi and oh-my-pi agents and also provides a standalone MCP (Model Context Protocol) server for use with clients like Claude Code, Cursor, and Windsurf.

Who it’s for

Developers and researchers who want a self-maintaining knowledge base that evolves with their projects, and those using AI agents that need a persistent, structured "working memory" for complex tasks.

Highlights

• Layered Recall: Simultaneously searches both a global personal vault (~/.llm-wiki/) and a project-specific vault.
• Open Knowledge Format (OKF) v0.2: Uses a portable, non-proprietary format to avoid vendor lock-in.
• Agent Working-Memory: Optional capability to record tool-call trajectories and distill them into reusable skill pages.
• Obsidian Compatibility: Knowledge is stored as Markdown with wikilinks, making it viewable in Obsidian.
• Automated Ingestion: Batch-processes raw sources into structured concept and analysis pages via wiki_ingest.
• MCP Server: Exposes wiki tools to any MCP-compatible AI client via stdio transport.