threat-vector-security/skills/preferences-memory

Preferences Memory

Use memory tools only when the user explicitly wants a stable fact or preference remembered beyond the current conversation.

Workflow

1. Confirm the thing is worth remembering.
   • stable preference
   • profile detail
   • standing instruction
   • recurring project context
2. Check for an existing related memory first with memory_search or memory_recall.
3. Save a concise normalized memory with memory_save.
4. Tell the user what was remembered in plain language.

Recurring Learnings

• If the user explicitly wants a lesson from this task retained for future conversations, normalize it into a short stable rule or project fact before saving it.
• Prefer one durable memory over a long incident log.
• Save only the part that will still matter later.

Do Not Turn This Into

• ad-hoc .learnings/ files
• automatic hook installation
• autonomous self-improvement loops
• silent edits to AGENTS.md, CLAUDE.md, or .github/copilot-instructions.md

Do Save

• preferred writing style or output format
• stable names, roles, or relationships the user wants reused
• recurring workflow preferences
• durable project context the user explicitly wants retained

Do Not Save

• one-off operational notes
• temporary status output
• approval bookkeeping
• transient paths, errors, or command results unless the user specifically asks to remember them

Gotchas

• Do not save memories implicitly from ordinary conversation; the user needs to want persistence.
• Do not store volatile logs, diagnostics, or ephemeral paths as durable memory.
• Do not turn memory into self-modifying prompts, local hook setup, or hidden configuration drift.