(\
       \\                             ____
        (\,--------------------------'()'------o
         (_    _________________________ /~"
          (_)_)                     (_)_)   woof.

Give your AI a memory.

Persistent memory, self-improving skills and a nightly routine for Claude Code and Codex CLI — made of nothing but files.

~ / anywhere
$ npx wienerdog@latest init

Prefer to look first? npx wienerdog@latest init --dry-run shows the plan and touches nothing.

No daemon · No server · No telemetry · MIT · uninstall removes every file

★ Star on GitHub

How it works

01

Interview

One command. Your AI interviews you and writes a real CLAUDE.md / AGENTS.md — who you are, how you work, what you care about.

02

One vault, both agents

A plain-markdown memory vault (Obsidian conventions, versioned in git) shared by Claude Code and Codex. Teach one, the other knows it too.

03

Dreaming

A nightly job reviews your sessions, promotes what matters into long-term memory, and turns repeated workflows into reusable skills. At most one git commit per night — everything is revertible.

Just files. That's the whole trick.

Other “personal AI” tools are apps — gateways, daemons, servers you must run and secure. Wienerdog is markdown, skills, and hooks interpreted by the AI tool you already pay for. Nothing listens, nothing phones home. It runs on your existing subscription, fully inside your provider's terms of service.

Honest about safety

Auto-written AI memory deserves skepticism. Nothing the nightly job proposes about you reaches a session until you approve the exact text. Read the threat model →

Most of this codebase is written by AI agents working from written specs behind a review gate — the repo dogfoods its own product.