akm

A universal, tool-agnostic knowledge system for AI agents. Plain Markdown, no database or server required. Works with Claude Code, Codex, OpenClaw, and any agent that can read files; optional zero-dependency scripts provide linting, operational enforcement, and export.

An LLM wiki gives an agent a map of knowledge. AKM gives it an operating system for knowledge: what to learn, where to store it, how to act on it, how to verify results, and where failures go so the next run is better.

Why

Agents forget. Plain RAG re-interprets the same documents every session, good syntheses evaporate when the session ends, and the same mistakes repeat. Compiled Markdown wikis (e.g. Karpathy's LLM Wiki pattern) fix the knowledge half — but knowing is not operating. AKM keeps the wiki pattern as its knowledge layer and adds operational memory, context, procedures, selective action logging, and an evaluation loop, so agents don't just know more — they work better over time.

The core idea is separation of stores. Long external knowledge, short operating rules, repeatable procedures, execution records, and verification criteria each fail differently when mixed. AKM stores them apart and reconnects them with wikilinks and frontmatter.

The seven layers

Plus two auxiliary layers: 80-outputs/ (deliverables) and 90-archive/ (deprecated notes).