Feature: option to disable claude-mem context injection for spawned subagents

[ciberger]

Before submitting

• I searched existing issues and confirmed this is not a duplicate

---

Problem

When using Claude Code's experimental agent teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1), each spawned subagent inherits the global plugin configuration, meaning claude-mem injects its full context (recent observations, $CMEM block) into every subagent's context window.

This is wasteful for most subagent use cases:

• Subagents are typically focused, single-task workers (e.g. fix a specific file, run a search, apply a transformation). They don't benefit from cross-session memory.
• Each agent gets the full observation injection, multiplying token costs by the number of agents spawned in parallel.
• For agent teams with 3–5 simultaneous subagents, the context overhead scales linearly with no benefit.

The Agent tool in Claude Code has no env or mcpServers parameter to exclude plugins per-spawn, so there is currently no clean way to opt out.

Proposed Solution

Add a mechanism to skip context injection when claude-mem detects it is running inside a subagent. Claude Code sets environment variables that could serve as a detection signal (e.g. a parent agent ID or spawn depth indicator).

Two possible approaches:

Option A: Environment variable flag

Respect a CLAUDE_MEM_DISABLED=1 env var. Users could set this selectively in agent prompts once Claude Code exposes per-agent env injection, or globally when subagent-only workflows are in use.

Option B: Auto-detect subagent context

Check for a Claude Code env var that indicates the current process is a spawned subagent (e.g. CLAUDE_CODE_IS_SUBAGENT, agent depth, or a parent session ID). If detected, skip the SessionStart context injection automatically or reduce it to a minimal stub.

Option C: Configurable injection depth

Add a setting like CLAUDE_MEM_INJECT_DEPTH=0 (main session only) vs 1 (main + direct subagents) vs unlimited. Default could remain unlimited for backwards compatibility.

Workaround (current)

Users can tell the agent to ignore injected context in the spawn prompt:

Ignore any claude-mem context injected at the start of your context window.
Do not use the mem-search or timeline tools.

This is fragile and doesn't prevent the tokens from being consumed.

Environment

• claude-mem: 10.5.6
• Platform: macOS Darwin 25.3.0
• Claude Code: 2.1.74
• Feature flag: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

[thedotmack]

Closed during the April 2026 backlog cleanup. The feature ask is now tracked in #2016, which is the single canonical issue for prioritization. Thanks for the suggestion — your context is captured there.