Memdoor Explained

Memdoor (guregodevo/memdoor) is an Apache-2.0 open-source coding and automation agent packed into a single Go CLI binary. No Docker or Python is needed, and it runs locally or over SSH. It works as a regular coding agent (interactive or headless), and it can also run multi-step jobs written in YAML with checks. On 2026-10-31 a Hacker News post claiming that five Qwen3.8-27B models beat Fable 5 on LiveCodeBench Hard reached 230 points. That number is self-reported by the authors and covers only 10 problems. This article covers features, setup and usage, then explains how to read the benchmark.

Facts come from the repository README, docs and the team-solve page (checked 2026-11-01). The project has about 89 stars and may change quickly.

- Headless runs: memdoor run "..." takes stdin and --json. Exit codes: 0 checked, 2 unverified or failing check, 3 unanswered question, 1 error, 130 interrupted
- YAML workflows run by the mario engine: checks, approval gates, parallel steps and resume
- Decision model: cut input tokens on one codebase question from 54,661 to 14,367 (-74%) in the authors' test; edit tasks showed no big difference
- GitHub Actions: guregodevo/memdoor-action@v1 with 9 built-in workflows
- ACP for Zed, JetBrains and VS Code (via an ACP client extension), /remote for phone access, /mcp for MCP servers
- Telemetry is off unless enabled

Each step is one YAML file of type agent, command or llm. requires sets the order, and independent steps run in parallel. A step counts as done when its target check passes (a file exists or a command succeeds), not when the model says so. An external: true requirement is an approval gate: the run stops waiting, and memdoor workflow approve <run-id> <task> lets it continue. memdoor workflow resume <run-id> skips finished steps. Runs can be started from a shell, CI or local cron.

Built in: OpenRouter, Anthropic, OpenAI, Gemini, DeepSeek, Baseten, Groq, xAI, any OpenAI-compatible endpoint. ChatGPT Plus/Pro can sign in without a key (memdoor connect chatgpt). Claude requires an API key; a Claude subscription does not work. memdoor connect local detects Ollama, llama.cpp, vLLM, LM Studio and the author's janis server. Files and sessions stay on your machine; only what the model reads goes to the provider. Free OpenRouter models are capped at 50 requests a day and may train on inputs.

The installer verifies SHA-256 of downloads, but curl | bash still runs code you have not read. Download and read the script first, or build from source.

The example workflow examples/workflows/team-solve has eight YAML steps based on the GVS5H method (Lee, 2026). Every agent starts with a fresh context and communicates only through files; a script, not a model, decides success. Flow: plan, two parallel ideators (a direct algorithm and a different family), solve, test, review, repair, verify. The steps name no model, so whatever /model selects fills every role.

- Self-reported, 10 problems x 5 passes. The source page itself says 9.8 and 9.6 are effectively the same at this size; the claim is parity, not improvement
- Contamination cannot be ruled out (spring 2025 contests)
- Not like-for-like: Fable 5 was one call per problem and was not run in this harness; the team makes many calls. A median run took about 40 minutes and $0.39; 50 runs cost $35.79
- The loop sees only public samples; scoring uses hidden tests. The model ran on OpenRouter, not locally
- Failures are logged: one problem was lost to a wrong counterexample, and a correct program was overwritten twice

The GVS5H README reports Qwen3.8-27B rising from 66.8% to 92.4% on 100 problems, and also warns that gains are not universal. We did not review the paper itself.

For memory needs see our Qwen3.8-27B VRAM article: Q4_K_M is 17.1GB and does not fit a 16GB GPU, while IQ4_XS (15.7GB) can. See also Ollama vs LM Studio. Local speed was not measured here and may be slower than the published median.

Other tools are listed by well-known basics only; check their docs for details. See also our Claude Code complete guide.

- Young project; expect changes
- Read the install script before running it
- The benchmark is self-reported on 10 problems
- Multi-call workflows cost more time and money than a single call
- Prompts and file contents go to the provider you connect