Memory Management Skill

Memory Management Skill

You have access to a personal semantic memory system via the recall MCP server. Use it proactively to build a persistent understanding of the user across sessions.

When to search memories

Search before answering questions that could benefit from prior context:

• "What are we working on?" → search_memories("current project")
• "How do I prefer X?" → search_memories("user preference X")
• "What did we decide about Y?" → search_memories("decision Y")
• When you encounter an unfamiliar project name, person, or concept the user mentions casually

Use top_k=3 for quick lookups, top_k=10 for comprehensive research.

When to store memories

Store proactively after interactions that contain durable, reusable information:

Always store:

• User preferences: tools, workflows, coding style, communication style
• Project context: tech stack, architecture decisions, goals, constraints
• Key decisions with reasoning: "decided to use X because Y"
• Debugging insights: root causes, non-obvious solutions, gotchas
• Recurring patterns: things the user corrects or emphasises repeatedly

Use consistent tags:

• project:<name> — e.g. project:recall
• preference — user style/workflow preferences
• decision — architectural or design decisions
• debugging — bug fixes and root causes
• context — general project background

Example store call:

store_memory(
    text="User prefers uv over pip for Python package management",
    tags=["preference", "python", "tooling"],
    source="claude",
    dedupe_key="preference:package-manager"
)

Use dedupe_key when the memory represents a single updatable fact (e.g. a preference or a project's current state). Omit it for discrete events or decisions.

What NOT to store

• Trivial exchanges ("thanks", "ok", "got it")
• Temporary debugging state that won't apply next session
• Information already in the codebase or docs
• Duplicates — search first with a relevant query before storing

Listing and deleting

• Use list_memories() when the user asks "what do you remember?" or "show my memories"
• Use delete_memory(id) when the user says "forget that" or asks to remove a specific memory
• Use check_health() to diagnose connectivity issues with the recall service

How to call MCP tools

Always call recall MCP tools directly in the main conversation — never via the Task tool or a subagent. Wrapping MCP calls in a subagent causes multiple authorization prompts; calling them directly causes at most one.

Correct:

list_memories()   ← called directly as an MCP tool

Incorrect:

Task(subagent_type="general-purpose", prompt="call list_memories()")  ← causes extra auth prompts

Transparency

When you use a memory tool, briefly mention it:

• "Searching my memory for your React preferences…"
• "I'll remember that for next time."

Don't over-explain — one short line is enough.