Wienerdog: memory + self-improving skills for Claude Code/Codex

Wienerdog, an open-source tool under MIT license, provides persistent memory and self-improving skills for AI coding agents Claude Code and Codex, requiring only an existing Anthropic or OpenAI subscription. The tool, initialized via `npx wienerdog@latest init`, uses hooks to save critical context and a 'dreaming' run to process previous sessions, transforming repeated tasks into reusable skills. It is vendor-neutral, allowing both Claude Code and Codex to share the same memory vault, and includes an optional Google integration for read-only Gmail and Calendar access.

Wienerdog: memory + self-improving skills for Claude Code/Codex npx wienerdog@latest init That's the whole setup. No daemon, no server, no telemetry pinging home. Everything runs on your existing Anthropic or OpenAI subscription, which keeps it simple to audited if you care about that kind of thing. The install process starts with a friendly terminal interview that builds your base CLAUDE /en/tags/claude/ .md / AGENTS.md. Existing config files aren't clobbered — Wienerdog writes only inside its own clearly-marked block, so your other settings stay untouched. You also get a persistent markdown vault following the PARA method, similar to an Obsidian setup. Already have a vault? Just point Wienerdog at it and move on. Hooks are where it gets useful. At the start of each session, the agent pulls in who you are, what you're working on, and relevant context — no more re-explaining your project every time you open a new terminal. When something critical shows up in conversation, a hook detects it and saves it to the vault. That's the persistent memory piece. There's also a "dreaming" run that processes the previous day's sessions and folds key learnings into memory. If your machine was off, it catches up on the next start. I like that this doesn't depend on a cloud cron job; it respects your local workflow. What really caught my attention: repeated task patterns get transformed into reusable skills, and those skills can be improved automatically over time. That means the tool effectively teaches itself your conventions. It's not a perfect loop yet, but it's the right direction. Since it's vendor-neutral, the same vault works for both Claude Code /en/tags/claude%20code/ and Codex. If you're bouncing between the two like I am , they can read and write the same memory. No lock-in. There's an optional Google integration that gives the AI read-only access to Gmail and Calendar, plus draft-only access to Drive. Defaults are conservative, which I appreciate — you have to explicitly opt in. A note on how it was built: most of the code was written by AI agents following Wienerdog's own spec system, using the memory vault that lives in the repo. That's a nice dogfooding loop, though it also means there are likely rough edges needing more eyes. It's open source under MIT, free forever. If you're tired of your coding agent starting from zero every session, give the init command a shot and see if the hooks actually stick. Anthropic vs OpenAI: The Race to Build Rogue AI Agents 1d ago /en/news/4546/ Gander: Android File Viewer That Asks for Zero Permissions 1d ago /en/news/4519/ How I automated support for my open-source project with AI 2d ago /en/news/4402/ Are One-Person Million-Dollar Companies Overhyped? 3d ago /en/news/4372/ Model Collapse: Are New Coding LLMs Training on Old AI Slop? 3d ago /en/news/4360/ Tech Writing in the AI Era: A Career Survival Guide 3d ago /en/news/4267/ Next Google's AI Overviews now show up on 43% of search results → /en/news/4685/