memory-project-context

name	memory-project-context
description	Provide first-turn understanding of the ai-memory repository with minimal tokens. Always use this skill when user asks what this project does, asks for implementation entrypoints, requests architecture explanation, or when a subagent needs a narrow read plan for this repository.
metadata	{"keywords":["ai-memory overview","project introduction","architecture entrypoints","first-turn understanding"],"triggers":["what is this project","introduce this repo","项目是做什么的","关键入口在哪里"]}

Memory Project Context

Use this skill as the first project-level semantic index before broad file reads.

Project Snapshot

• Project: ai-memory
• Primary goal: persistent developer memory layer for AI coding tools
• Core model: dynamic runtime memory over static snapshots
• Data source of truth: ~/.memory/profile.md, preferences.md, infra.md, projects/*.md

Fast Answers

When user asks for a quick intro, answer with:

1. One sentence on project goal.
2. One sentence on architecture split: ai-memory CLI, memory files, injection layer.
3. One sentence on why dynamic context is preferred.

Entrypoints

• ai-memory: command router, context rendering, inject modes
• lib/learn.sh: memory update pipeline from LLM output
• docs/core-idea.md: architecture principles
• note.md: implementation progress and roadmap
• .github/copilot-instructions.md: Copilot routing instructions generated by inject

Routing Rules

1. If task is architecture or design rationale:

• Read docs/core-idea.md then ai-memory.

2. If task is command behavior or bug fix:

• Read ai-memory first, then specific module files.

3. If task is memory update flow:

• Read lib/learn.sh and related helper functions in ai-memory.

4. If task is token/context efficiency:

• Use .github/skills/memory-token-audit/SKILL.md.

Minimal Read Plan For Subagents

• Start with only one file section matching the task.
• Expand to a second file only if the first cannot answer.
• Avoid loading full generated instruction snapshots unless user asks for full dump.

Output Contract

• Start with direct answer.
• Include only required files and commands.
• Explicitly call out whether recommendation changes token cost or not.