#Plugin context pollution:

1 messages · Page 1 of 1 (latest)

heady reef
#

I am on version 2026.4.20;
When OpenClaw loads stock plugins (provider adapters, disabled channel plugins, etc.) — do any of them inject content into the agent's context window, or is context injection limited to memory plugins and skills? Trying to understand if 62 loaded plugins is a context concern or purely a startup/RAM concern.

calm zenithBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

fallow lilyBOT
fleet quiverBOT
#

Mostly startup/RAM/import overhead, not automatic prompt bloat.

Short version for 2026.4.20:

  • No, stock loaded plugins do not all inject into the context window.
  • Yes, context injection is broader than just memory plugins + skills.

What can affect prompt/context:

  • Skills: eligible skills are injected as a skills list.
  • Active memory plugin: memory guidance / recall prompt path.
  • Active context engine: can add systemPromptAddition.
  • Active provider plugin: can add provider-specific system prompt guidance.
  • Enabled plugins with prompt hooks: before_prompt_build or legacy before_agent_start.
  • Plugins that add tools: their tool schemas/list count toward context if those tools are available.

What generally does not affect prompt/context: