ai-agent-memory

A practical, git-native implementation of the Karpathy LLM Wiki pattern for persistent coding-agent memory.

The problem: AI coding agents are stateless — every session starts from scratch. But it gets worse: if you use GitHub Copilot with multiple agents (built-in + 3rd-party like Claude, Codex), each has its own instruction format. Maintaining separate config files across every project, for every agent, doesn't scale. And knowledge stays siloed per project — a pattern you discovered in one client project is invisible to the next.

The solution: A central markdown wiki that all your agents share, across all your projects. Each project points to it through a single AGENT.md; each agent-specific config file (.github/copilot-instructions.md, CLAUDE.md, etc.) simply redirects there. The wiki grows with every session — call end-session and the agent captures lessons, decisions, and patterns back into the wiki. Knowledge learned in one project (a deployment gotcha, an API pattern, a debugging technique) is immediately available in every other project. The next session, with any agent, on any project, starts with all that accumulated knowledge.

Works with: GitHub Copilot · Claude Code · Codex · Cursor · Gemini CLI · Aider

Every session feeds the wiki, and the wiki feeds every session — the Karpathy loop at a glance: