Imported from MTEnt/grok-agents-memory (AGENTS.md). Install upstream with npx skills add MTEnt/grok-agents-memory. Copyright stays with the author.

Grok Agent Workspace Rules

This project is a Grok agent bootstrap workspace with a real SQLite-backed memory layer exposed through the project MCP server named agent_memory.

Memory Contract

• Use the agent_memory MCP tools for durable agent and project memory.
• Do not claim a memory was saved unless agent_memory__memory_write or the equivalent use_tool call succeeds.
• Do not simulate persistence in prose. If the MCP server is unavailable, say SQLite memory is not connected and stop before pretending to remember.
• At the start of meaningful work, call context_pack_get for the active agent. This is the canonical startup context: agent profile, project, memories, and job board.
• At handoff or completion, write durable decisions, user preferences, resolved bugs, implementation outcomes, and agent-specific history back to SQLite.
• Keep memory entries short, factual, and searchable. Use tags.

Job Board Contract

The SQLite job board is the source of truth for current work.

• Do not start substantial work until a job exists.
• Claim a job with job_claim before doing substantial work.
• Append progress, blockers, review notes, and handoffs with job_append_event or job_report.
• Finish with job_complete or job_block; do not leave real work silently in progress.
• Use context_memory_ids on jobs to connect work to the project brief and other durable memories.
• If no relevant job exists, ask project-founder to create one or create a narrowly scoped job before continuing.

Empty Project Intake

When the working folder has no product/code files yet, the project-founder agent must question the user before scaffolding. Ignore this toolkit's bootstrap files when deciding emptiness: .git, .grok, mcp, scripts, tests, AGENTS.md, README.md, .gitignore, and generated cache/database files. The first response should be a short interview, not file creation or architecture speculation.

Ask at most six questions in one message:

1. What are we building, and who is it for?
2. What is the first workflow that must be usable?
3. Is this a web app, mobile app, CLI/tool, API/service, library, game, or other?
4. Any preferred stack, constraints, integrations, or deployment target?
5. What should the experience feel like?
6. Which agent team should be active: lean, standard, or full?

After intake, write a project brief to SQLite and update the project record. Then create job-board items for the selected specialist team. Do not spawn implementation work before the brief and initial jobs exist in SQLite.

Agent Team

Use persistent named agents rather than disposable role-play:

• project-founder: intake and orchestration
• product-lead: product scope, workflows, and acceptance criteria
• architect: technical design, risks, and boundaries
• implementer: implementation and verification
• reviewer: adversarial review
• memory-curator: memory cleanup and consolidation

Each named agent must load its context pack before acting, claim/report/complete or block its job, and append a short durable memory only when the result should survive beyond the current job.