Bug

The active-memory plugin passes the full agent system prompt (skills, tools, behavioral instructions, SOUL, etc.) to the memory search model (gemma-4-31b-it). This is unnecessary — the memory search agent only needs the recent conversation context, not the full agent configuration.

Root cause

The memory search prompt appears to include the entire agent bootstrap context:

• System prompt with tools, skills, instructions
• SOUL, AGENTS.md, identity configuration
• All active skill declarations

When Gemma only needs: "here is the recent conversation, return relevant memories."

Impact

• Token bloat: Full agent context is often 5-10x larger than needed
• Slower generation: More tokens to process = slower response from Gemma
• Higher cost: Unnecessary token usage on every active-memory call
• Timeout risk: Larger prompts take longer to process, increasing chance of 15s timeout

Expected behavior

The memory search agent should receive a lean prompt containing only:

• Its role as memory search agent
• The recent conversation context (or relevant excerpt)
• Instructions to return only relevant memories

The full agent system prompt, skill declarations, tool listings, and behavioral instructions are not needed for memory search and should be stripped.

Environment

• OpenClaw v2026.4.12
• active-memory plugin: enabled
• Model: openrouter/google/gemma-4-31b-it
• active-memory promptStyle: balanced (default)