[FEATURE] Option to disable context injection for spawned subagents

[github-actions[bot]]

Originally posted as issue #2016 by @thedotmack
#2016

Type: feature / enhancement request

Description: Option to disable context injection for spawned subagents

---

Consolidated during the April 2026 backlog cleanup. Original: #1464

[AVXLab-Studio]

Confirming I ran into practically the same need described in the original request (#1464 / #2016) — wanting to disable claude-mem specifically for spawned/headless workers, without losing anything else.

Adding some data in case it helps move this forward. --settings '{"enabledPlugins":{"claude-mem@thedotmack":false}}' already works today as a per-invocation override to fully suppress claude-mem for a single claude -p call, without touching global config. Verified empirically (v13.8.0):

• SessionStart's additionalContext injection (the "recent context" dump) is fully suppressed
• All 21 mcp__plugin_claude-mem_mcp-search__* tool schemas drop out of the tool listing
• All 16 claude-mem:* skills drop out of the skill listing
• Native CLAUDE.md / project memory (separate system) stays completely unaffected — confirmed the worker still correctly recalled CLAUDE.md content with the flag applied
• ~28.6K fewer fresh-context tokens on a single trivial haiku turn (would compound more on longer workers, since the 21 tool schemas resend every turn)

This seems to be exactly what this thread is asking for — just undocumented, so it takes reverse-engineering to find. Might be worth documenting it as the official per-spawn workaround, or promoting it to a first-class flag like --no-claude-mem for discoverability, since right now it only works if you already know the exact plugin ID string.

Hope this helps close the gap — really appreciate the tool and want to see it get even leaner for this use case.