| name | ghostcrab-memory |
| description | Use when an agent has access to GhostCrab MCP and needs durable project memory, onboarding capture, context lookup, relationship tracking, or handoff continuity across sessions. Adapt this template to the target agent's skill, rule, or profile format. |
GhostCrab Memory
Purpose
Use GhostCrab as durable working memory for projects, decisions, facts, relationships, schemas, workflows, and handoffs. Prefer it when context should survive the current chat or be discoverable by another agent later.
Before Work
- Check that the GhostCrab MCP server is connected.
- Identify the active project or workspace.
- Search existing GhostCrab memory before adding new durable facts.
- Treat retrieved memory as context, not proof. Verify drift-prone facts from the current repo, runtime, or source of truth.
What To Store
Store concise, reusable facts:
- project conventions and commands
- durable architectural decisions
- validated setup steps
- known failure modes and fixes
- handoff summaries with next concrete action
- domain vocabulary, schemas, and relationships
Do not store:
- secrets, tokens, passwords, private keys
- unverified speculation
- large logs or full source files
- transient todo items that will be obsolete by the next session
- duplicate facts already present in memory
Working Pattern
When starting a task:
- Search by project name, repo path, component, and user wording.
- Read only the memory entries that are directly relevant.
- Use current local evidence to confirm anything likely to drift.
- Continue the work from the current checkout, not from memory alone.
When closing a task:
- Record what changed.
- Record what was validated.
- Record what remains fragile or unverified.
- Record the next concrete step if the work is not complete.
Query Hints
Useful search keywords:
- repo path
- package name
- command name
- feature name
- error text
- agent/client name
- exact file path
MCP Command Baseline
GhostCrab should be launched as:
gcp brain up
Portable no-install form:
npx -y --package=@mindflight/ghostcrab-personal-mcp@latest gcp brain up
If the client has a limited PATH, prefer:
node /absolute/path/to/node_modules/@mindflight/ghostcrab-personal-mcp/bin/gcp.mjs brain up
Adaptation Notes
For a new agent:
- Keep this
SKILL.md body short.
- Add only the loader metadata required by the target agent.
- Keep MCP config outside the skill unless the target agent expects it inside the skill folder.
- Preserve the "search before store" and "verify drift-prone facts" rules.
