AgentMemory for Codex on Windows — source-only preview with project-scoped recall

[M-T-D-N]

I’m a Windows user in South Korea, not a conventional software developer, and I use Codex to turn requirements and acceptance checks into working tools.

I published AgentMemory for Codex on Windows, an independent, source-only downstream of AgentMemory v0.9.29.

I made it because I believe durable, project-scoped memory can make Codex more useful for long-running work. A later task should be able to recover relevant context without mixing unrelated projects or turning cross-project recall into unrestricted writes.

The supported Windows profile currently:

• captures normal main-agent prompts and final responses through four managed Codex hooks: SessionStart, UserPromptSubmit, Stop, and SessionEnd;
• keeps writes, deletion, curation, and provenance exact-project scoped;
• supports bounded, source-labelled federated recall;
• uses an authenticated loopback MCP endpoint, with optional credential-free loopback Qwen graph extraction and no external model fallback;
• provides documentation in English, Korean, and Japanese.

The attached 40-second video is an assembled walkthrough based on a verified live save → recall run, not a continuous recording. Internal identifiers were redacted.

Current limits: this first preview has no prebuilt or signed installer and must be built from source on Windows. It is not an official upstream release. Most downstream modifications were generated and revised by OpenAI Codex from my requirements. I validated the result with automated tests and live Windows/Codex use, but I did not manually review the source, and no independent audit has been performed.

Release: v0.1.0-preview.1

Sharing this where it may be useful to Windows/Codex users who are comfortable evaluating a source preview.

agentmemory-demo-candidate-v3.mp4