openclaw - 💡(How to fix) Fix Active Memory sub-agent receives all tools instead of toolsAllow filter (toolsAllow not forwarded in runEmbeddedPiAgent) [1 comments, 2 participants]

Bug Description

Active Memory sub-agent is supposed to only have access to memory_search and memory_get (via toolsAllow), but the toolsAllow parameter is not forwarded through runEmbeddedPiAgent → runEmbeddedAttemptWithBackend. As a result, the sub-agent receives all configured tools and never calls the restricted memory tools.

Reproduction

Environment

• OpenClaw version: 2026.4.20
• Active Memory plugin: enabled, queryMode: "recent", promptStyle: "recall-heavy"
• Models tested: glm-5.1, grok-4.20-fast (via proxy providers)

Steps

1. Configure active-memory plugin with any model
2. Ask a question that should trigger memory recall (e.g., "what is my favorite singer")
3. Enable persistTranscripts: true for debugging

Observed Behavior

With glm-5.1:

• Sub-agent receives all tools, ignores memory_search
• Fabricates answer without calling any tool: "User's favorite singer is 周杰伦."
• Answer is incorrect (fabricated)

With grok-4.20-fast:

• Sub-agent receives all tools, ignores memory_search
• Attempts to use web_search instead (wrong tool)
• Returns NONE

Neither model calls memory_search or memory_get, despite the plugin explicitly setting toolsAllow: ["memory_search", "memory_get"].

Transcript Evidence

The sub-agent transcript shows:

• No memory_search or memory_get tool calls
• Model thinking includes references to web_search or direct fabrication
• Only a user message → assistant text response (no tool invocation)

Root Cause

Same as #66581 — runEmbeddedPiAgent accepts toolsAllow but does not forward it to runEmbeddedAttemptWithBackend. The sub-agent gets all 60+ tools instead of the restricted 2, causing:

1. Tool confusion: models pick wrong tools (web_search) or skip tools entirely
2. Context bloat: full tool schemas inflate the prompt to ~37k tokens
3. Timeout: large prompts cause slow responses and frequent timeouts

Related Issues

• #66581 — Original report of this exact bug (toolsAllow not forwarded)
• #65998 — MCP tool schemas bypass toolsAllow filter
• #66157 — Active-memory timeout with queryMode="message" (likely caused by this bug)

Workaround

None effective. The bug is in the core runner, not in plugin config. Setting persistTranscripts: true helps with diagnosis but does not fix the tool injection.