remember-people

Remember people

Use the people-memory MCP as the durable human memory behind the conversation.

On every name mention

1. Call search_people before answering.
2. If one record matches, use it silently as context.
3. If several records match and the answer depends on identity, ask which person the user means.
4. If no record matches, continue the conversation. Create the person when the user provides a durable fact or explicitly says they know them.

Do not ask “who is that?” until the graph search has failed.

Save durable facts

Write facts that should survive the current chat, including:

• job, organization, role, city, country, birthday, family, preferences, goals, or constraints;
• how two people know each other;
• who introduced whom;
• a dated call, meeting, message, coffee, meal, or visit;
• a new email, phone, LinkedIn URL, or other stable identifier.

Use remember_person for core fields and one simple fact. Use add_fact, record_interaction, and connect_people for their specific records. Save the source and mark deductions as inferred.

Do not store passing jokes, guesses, judgments, full message bodies, or facts about the user's inner life. Keep compact facts that help a future conversation.

Resolve before writing

Resolve by email or phone, then exact name. If a tool returns needs_confirmation, ask the user and retry with confirmed_new=true or overwrite=true only after they decide. Never silently merge similar names. Never replace a user-stated value with imported or inferred data.

Answer network questions

• Use get_person for “who is X?”
• Use search_people for organization, role, city, or free-text questions.
• Use find_intro_path for warm introductions.
• Use stale_contacts for neglected relationships.
• Use read_query for advanced filters that semantic tools do not cover.

Treat returned records as private third-party data. Do not quote or export them to anyone except the user who owns the graph.