| name | memory-router |
| description | Route tasks in the ai-memory project and load only the minimum required memory context. Always use this skill when user asks about ai-memory architecture, inject modes, copilot-instructions redesign, project memory lookup, learn pipeline behavior, or when a subagent should avoid broad file reads. |
| metadata | {"keywords":["routing","minimal context","inject modes","copilot instructions"],"triggers":["how should we route this task","inject strategy","copilot instructions redesign","最小化读取路径"]} |
Memory Router
Use this skill to avoid overloading context while keeping project understanding accurate.
Goal
- Identify task type quickly.
- Load minimum context first.
- Escalate to deeper files only when needed.
Task Classification
- First-turn project understanding:
- Read .github/skills/memory-project-context/SKILL.md first.
- Inject and instruction design:
- Read ai-memory and .github/copilot-instructions.md first.
- Core architecture:
- Read docs/core-idea.md first, then ai-memory.
- Learn pipeline:
- Read lib/learn.sh and ai-memory config/inject branches.
- Project status:
Retrieval Ladder
- ai-memory capability
- .github/skills/memory-project-context/SKILL.md (project semantic index)
- ai-memory context
- ai-memory project (only if project-specific)
- Open exact file sections for implementation details
Response Contract
- Start with direct answer in 1-3 lines.
- Then list only the files actually needed.
- For subagent tasks, state a concrete read plan before opening files.
- If proposing edits, keep diff surface small and mention validation steps.
Token Discipline
- Do not paste large snapshots unless asked.
- Prefer summaries over full file dumps.
- For VS Code instructions, keep the main file lightweight and route to project summary/skills.
