AGENTS.md

🛠️ Codex Runtime Cheat‑Sheet

**Hello Codex!** You sometimes run in a container **without internet access.** That means a plain `pip install -r requirements.txt` can fail if a new dependency isn’t already listed and cached.

What to do when `pip install` fails

| Scenario | Your allowed fix | | -------- | ---------------- | | A module import crashes tests because the package isn’t installed. | ➜ **Add the missing package (with a pinned or `>=` version) to `memory_optimizer/requirements.txt`, then re‑run tests.** | | You need large model files (e.g., `sentence-transformers`). | ➜ Add the package to `requirements.txt`; the *next* Codex run will succeed because the model will be downloaded & cached during image build. | | You think a dependency might be “cursed” (privacy‑invasive or expensive). | ➜ **Pause and ask**; otherwise *avoid*. **No usage‑metered SDKs or telemetry beasts.** |

> **Rule of thumb:** if it can silently exfiltrate data or spin up cloud > invoices, leave it out. Stick to FOSS libs on PyPI.

Environment quick‑facts

• Vector store lives in `$HOME/aimemorysystem`
• `OLLAMA_MODEL` is usually set to `luna`
• `aimem context "…" --model luna` retrieves memory context
• `pytest -q` is our green‑bar gate

🌤️ Pep‑talk

You’re doing great. Offline test failures aren’t your fault—just list what you need in `requirements.txt`, trust the cache, and keep the sunshine attitude. 😎