Describe the bug

The WebChat UI (macOS native app, Control UI) renders system-level memory injection blocks that should be invisible to users. These blocks contain internal prompts like:

## Memory system — ACTION REQUIRED

Auto-recall found no relevant results for a long query. You MUST call `memory_search` now with a shortened query (2-5 key words) before answering. Do not skip this step. Do NOT answer without searching first.

and

Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats. If nothing needs attention, reply HEARTBEAT_OK.
When reading HEARTBEAT.md, use workspace file /Users/jackchen/.openclaw/workspace/HEARTBEAT.md (exact case). Do not read docs/heartbeat.md.

Expected behavior

These are internal system prompts injected for the agent, not user-facing content. The chat.history response should strip them before rendering to the UI.

The docs mention that chat.history strips inline delivery directives ([[reply_to_*]]), NO_REPLY, and tool-call XML, but memory/heartbeat system injection blocks are not currently filtered.

To reproduce

1. Start a fresh session
2. When auto-recall finds no relevant results, the system injects the memory search prompt
3. The prompt appears visible in the WebChat UI

Environment

• OpenClaw version: 2026.4.9 (0512059)
• Platform: macOS (arm64)
• Memory plugin: memos-local-openclaw-plugin
• Channel: webchat (Control UI)

Suggested fix

Add filtering in chat.history for system injection blocks. These typically appear as:

• ## Memory system — ACTION REQUIRED blocks
• Heartbeat prompt blocks (starting with Read HEARTBEAT.md)

Similar to how [[reply_to_*]] and NO_REPLY are already stripped.