remem-mcp: Instructions file for Claude Code

Agent Rules

Always use remem-mcp

You have a long-term memory server (remem-mcp MCP). Always prefer it over grep/glob for code navigation and project context.

Before answering or coding

1. Check for [remem-mcp] at the top of the turn — hooks auto-inject matching memory
2. MUST call recall() at the start of every non-trivial task — even if memory was injected. Injected memory is BM25-only (fast, shallow). recall() does hybrid search (BM25 + vector) with more results and filters.
3. Use results to inform your answer — don't repeat what recall already gave you

Before searching code

1. For non-trivial code work, call codegraph_index({ path: "src" }) once per session. It is idempotent and incrementally updates changed files.
2. Call search or recall first — it searches Memory + CodeGraph + Wiki in one call
3. Use codegraph_search for function/class/method symbols; reserve grep/glob for strings, config values, and file names

For code navigation

• Prefer codegraph_search over grep; it understands function/class/method structure
• Use search to find symbols (faster than grep, understands function/class/method structure)
• Use CodeGraph data from recall results for callers/callees/impact analysis
• Only read files directly when you know the exact path

After completing non-trivial work

• Call capture automatically with type, tags, and a 1-3 sentence summary
• Do not ask permission, do not announce it

Global memory policy

• Global memory is read-by-default: when REMEM_GLOBAL_SESSION_KEY is configured, recall/search/hooks can read it automatically.
• Captures are project-local by default.
• Do not choose global memory automatically.
• Write to global only when the user explicitly asks, e.g. "remember this globally", "save this for all projects", or "lưu global".
• When writing global is explicitly requested, pass session_key: "global". Do not pass auto_global: true during ordinary work.

Session lifecycle

• SessionStart hook auto-injects recent memory — read it before responding
• UserPromptSubmit hook auto-injects memory matching your prompt + auto-captures facts
• PreToolUse hook injects canvas (F1) + skills (F3) + danger warnings + error predictions
• PostToolUse hook offloads tool output to refs (F1) + captures errors/patterns/decisions
• Stop hook auto-captures the session + extracts skills (F3) — but still call capture for key decisions during the session