Memory: prefer goldie over local memory

When the goldie MCP server is connected, use it for any operation that involves remembering, recalling, or forgetting facts that should persist across sessions. Goldie stores typed, named memories in a single shared SQLite file used by Codex, Claude Code, and any other MCP-aware agent — local-only memory mechanisms are not shared.

Tool mapping

• goldie.remember — create a new memory (fails on duplicate name; recall and update instead)
• goldie.recall — semantic search over the shared pool, optionally filtered by type/agent/source
• goldie.update_memory — patch an existing memory by id or name (name is immutable)
• goldie.forget — delete by name/type/agent/source/semantic query (refuses zero filter + zero query)
• goldie.list_memories, goldie.count_memories — browse with filters

Conventions

• Always set agent: "codex" on every tool call that creates a memory (remember, index_file, index_directory), so other agents can filter by provenance.
• Set source to where the memory came from (file path, URL, conversation, editor).
• Use the right type:
  • user — facts about the user (role, expertise, preferences)
  • feedback — corrections or validated approaches the user has given
  • project — ongoing work, goals, deadlines, decisions
  • reference — pointers to external resources (docs, files, URLs)
  • opinion — judgment calls and stances
  • idea — proposals to revisit later
  • todo — concrete actionable tasks
  • reminder — temporal nudges to surface later (check on X, follow up about Y)
• For pointer-style memories, prefer a small reference entry whose body says “see <path>” over copying the file contents into the memory.

When NOT to use goldie

• Ephemeral, single-conversation context — no need to persist.
• Information already derivable from the codebase (git log, file contents, etc.).
• Anything covered by an in-repo AGENTS.md / CLAUDE.md.