### Summary claude-mem ships 30 skills, and all of them are model-invocable. Many users need only a few. For me, the one I actually need is `mem-search`. There's currently no claude-mem setting to turn the others off, and Claude Code's own `skillOverrides` is ignored for third-party plugin skills (anthropics/claude-code#47747, re-filed as https://github.com/anthropics/claude-code/issues/105570). So users can't trim the list. ### Why it matters - Every skill description is loaded into the listing every session (~1.6k tokens for the 30). - Several skills overlap with Claude Code built-ins or common plugins. Because they're model-invocable, the model can pick them over the user's own workflow: `make-plan`/`do` overlap plan mode, `smart-explore` the Explore agent, `handoff`, `babysit`, `release`/`version-bump`. - The only workaround is `permissions.deny: ["Skill(claude-mem: )"]`. That blocks invocation but leaves the skill in the listing. ### Possible approaches (any one would help) 1. A setting such as `CLAUDE_MEM_SKILLS_ENABLED` (allow-list) or `CLAUDE_MEM_SKILLS_DISABLED` (deny-list), applied when the plugin loads. 2. Ship only the core memory skills (`mem-search`, `how-it-works`) by default, and move the workflow skills (build, release, babysit, standup, design, decks…) into an optional companion plugin. 3. Mark workflow skills `disable-model-invocation: true`, so they stay available as slash commands without competing for automatic invocation or listing space. Thanks for claude-mem. The per-feature switches for context injection (`CLAUDE_MEM_CONTEXT_OBSERVATIONS`, `CLAUDE_MEM_DISABLE_FILE_CONTEXT`, …) made it easy to run it query-only. The same kind of control for skills would round it out.
