The agent's guide

This is the manual an agent (or the person who runs one) reads: how memprove is installed, how a repository is enabled, what one Session does from begin to end, what the Context Pack says, how the Claude Code hook and the MCP server work, how a Session another agent left is taken over, and how the record is corrected. memprove --help and memprove <command> --help print the usage and exit codes of every command without a daemon; this guide says the same as they do, and the tests hold the two together. docs/architecture.md says how it is built; docs/takeover-eval.md says how the takeover is evaluated with a real client.

Terms. A Project is a Git repository the operator registered; a Workspace is one of its worktrees. A Task is a unit of work; a Session is one agent’s stretch of work on it, with a declared agent (name, role, parent Session). A run (memprove run) records a command’s result and seals a Snapshot of the repository state; a fact cites runs as its Evidence. A Checkpoint records facts, open items, decisions and the one next action in force. The Capsule is the Task’s record as the daemon judges it now; the Pack is its rendering within a budget. A fact’s verdict is verified, stale, unconfirmed or tampered, by deterministic rules, never by a model.

Sections

• Install, upgrade and backup
• Enabling memprove in a project
• Unattended takeover
• One Session, start to end
• Several runs in one state
• When a test fails
• Correcting the record: supersede, withdraw, reaffirm, decide
• A Session that was not ended
• The Context Pack (resume –pack)
• The SessionStart hook (memprove hook)
• The MCP server (memprove mcp)
• Completing a Task
• memprove run exit codes
• Exit codes of the other commands
• Known limits

From docs/guide.md at 2625064cbf0a · Markdown