Fully local -- embeds the text with a local model and writes it to a local
SQLite file. No LLM call, no network. Use this whenever the user tells you
something worth remembering across sessions (preferences, facts, decisions).

Args:
    content: The text to remember (a fact, preference, or note).
    user_id: Namespace to store under (default "default"). Use per-end-user.

Read-only, fully local semantic search. Call this before answering when the
user refers to past context, preferences, or previously-shared facts.

Args:
    query: What to look for (a question or topic).
    limit: Max memories to return (1-50, default 5).
    user_id: Namespace to search (default "default").

Destructive. Finds the best-matching memory and removes it. If nothing
matches, nothing is deleted.

Args:
    query: Describes the memory to remove.
    user_id: Namespace to delete from (default "default").

Args:
    user_id: Namespace to clear (default "default"). This never clears other
        users' namespaces.