llm-wiki-vault

A persistent, compounding knowledge base for AI agents — designed to replace brittle RAG pipelines with a structured, git-tracked markdown vault that agents read, write, and maintain across sessions.

What this is

Most agent memory systems are either stateless (the agent forgets everything between sessions) or over-engineered (vector databases, embeddings, complex retrieval stacks). This vault takes a different approach: plain markdown files, a strict directory convention, and agents that are trained — via AGENTS.md — to compile, query, and maintain knowledge deterministically.

The result is a knowledge base that compounds over time. Every session adds to it. Every source gets processed once and stays compiled. Agents never re-read raw sources; they read wiki pages.

How to use it

If you are a human, start with:

• IDEA.md — define what this vault is for and what's out of scope
• LLM-WIKI-ARCHITECTURE.md (in outputs/) — full design rationale and directory map
• tools/wikictl — bash CLI for day-to-day vault operations (status, bridge, lint, log)

If you are an agent, start with:

• AGENTS.md — your complete operational schema: session protocols, all workflows, hard rules, token budgets

Key design principles

Compile-first, never query raw. Sources are ingested once into wiki pages. Agents read wiki, not raw/.

Persistent compounding. Every session's work stays. The vault gets smarter over time rather than resetting.