# Agent Restart Protocol

Session memory does not survive a restart. The reliable technique is:

1. keep a repo-local context file
2. tell the restarted agent exactly what to read first
3. force it to inspect git state before it acts

## Copy-Paste Prompt For QuantumLearning

Use this when restarting an agent inside `QuantumLearning`:

```text
Read AGENT_CONTEXT.md first. Then read README.md, OPERATIONS.md, FIRST_RUN.md, configs/curriculum.toml, configs/mastery_blueprint.toml, and tests/test_course_flow.py. After that, inspect `git status --short --branch` and `git log --oneline -5`, summarize the current project state and constraints, and only then continue work.
```

## Copy-Paste Prompt For A New Coursework Repo

Use this as the starting pattern for a new repo like `NTT-learning`:

```text
You are starting a new local-first Jupyter learning platform. Before building, inspect the repo state, create a repo-local `.venv`, and establish a single supported end-to-end walkthrough. The notebooks must distinguish clearly between META cells, MANDATORY cells, and FACULTATIVE cells. Difficulty 1-3 is reserved for mandatory cells; 4-10 is reserved for facultative cells. Hide plumbing where possible, keep the route unambiguous, and maintain repo-local operational scripts so a normal user can start, stop, restart, validate, and monitor the platform without agent help. After inspection, summarize the immediate setup plan and then execute it.
```

## Technique Summary

The agent cannot truly "remember" across restarts unless the memory is written into the repo. The durable method is:

- repo-local context file
- repo-local operational docs
- explicit restart prompt
- git inspection before action