memory-neo4j: strip channel metadata wrappers, reject system infra messages in attention gate

extensions/memory-neo4j/extractor.test.ts

@@ -82,6 +82,46 @@ describe("passesAttentionGate", () => {
     );
   });
 
+  it("should reject system infrastructure messages", () => {
+    // Heartbeat prompts
+    expect(
+      passesAttentionGate(
+        "Read HEARTBEAT.md if it exists (workspace context). Follow it strictly.",
+      ),
+    ).toBe(false);
+
+    // Pre-compaction flush
+    expect(passesAttentionGate("Pre-compaction memory flush. Store durable memories now.")).toBe(
+      false,
+    );
+
+    // System cron/exec messages
+    expect(
+      passesAttentionGate(
+        "System: [2026-02-06 10:25:00 UTC] Reminder: Check if wa-group-monitor updated",
+      ),
+    ).toBe(false);
+
+    // Cron job wrappers
+    expect(
+      passesAttentionGate(
+        "[cron:720b01aa-03d1-4888-a2d4-0f0a9e0d7b6c Memory Sleep Cycle] Run the sleep cycle",
+      ),
+    ).toBe(false);
+
+    // Gateway restart payloads
+    expect(passesAttentionGate('GatewayRestart:\n{ "kind": "restart", "status": "ok" }')).toBe(
+      false,
+    );
+
+    // Background task completion
+    expect(
+      passesAttentionGate(
+        "[Sat 2026-02-07 01:02 GMT+8] A background task just completed successfully.",
+      ),
+    ).toBe(false);
+  });
+
   // --- Should ACCEPT ---
 
   it("should accept substantive messages with enough words", () => {
@@ -248,6 +288,55 @@ describe("extractUserMessages", () => {
     const result = extractUserMessages(messages as unknown[]);
     expect(result).toEqual([]);
   });
+
+  it("should strip Telegram channel metadata and extract raw user text", () => {
+    const messages = [
+      {
+        role: "user",
+        content:
+          "[Telegram Tarun (@ts1974_001) id:878224171 +1m 2026-02-06 23:18 GMT+8] I restarted the gateway but it still shows UTC time\n[message_id: 6363]",
+      },
+    ];
+    const result = extractUserMessages(messages);
+    expect(result).toEqual(["I restarted the gateway but it still shows UTC time"]);
+  });
+
+  it("should strip Telegram wrapper and filter if remaining text is too short", () => {
+    const messages = [
+      {
+        role: "user",
+        content:
+          "[Telegram Tarun (@ts1974_001) id:878224171 +1m 2026-02-06 13:32 UTC] Hi\n[message_id: 6302]",
+      },
+    ];
+    const result = extractUserMessages(messages);
+    // "Hi" is < 10 chars after stripping — should be filtered out
+    expect(result).toEqual([]);
+  });
+
+  it("should strip media attachment preamble and keep user text", () => {
+    const messages = [
+      {
+        role: "user",
+        content:
+          "[media attached: /path/to/file.jpg (image/jpeg) | /path/to/file.jpg]\nTo send an image back, prefer the message tool.\n[Telegram Tarun (@ts1974_001) id:878224171 +5m 2026-02-06 14:01 UTC] My claim for the business expense\n[message_id: 6334]",
+      },
+    ];
+    const result = extractUserMessages(messages);
+    expect(result).toEqual(["My claim for the business expense"]);
+  });
+
+  it("should strip System exec output prefixes", () => {
+    const messages = [
+      {
+        role: "user",
+        content:
+          "System: [2026-01-31 05:44:57 UTC] Exec completed (gentle-s, code 0)\n\n[Telegram User id:123 +1m 2026-01-31 05:46 UTC] I want 4k imax copy of Interstellar\n[message_id: 2098]",
+      },
+    ];
+    const result = extractUserMessages(messages);
+    expect(result).toEqual(["I want 4k imax copy of Interstellar"]);
+  });
 });
 
 // ============================================================================

extensions/memory-neo4j/extractor.ts

@@ -872,15 +872,24 @@ export function extractUserMessages(messages: unknown[]): string[] {
     }
   }
 
-  // Strip injected context blocks (auto-recall prepends these into user messages)
-  // then filter out noise
+  // Strip injected context, channel metadata wrappers, and system prefixes
+  // so the attention gate sees only the raw user text.
   return texts
-    .map((t) =>
-      t
-        .replace(/<relevant-memories>[\s\S]*?<\/relevant-memories>\s*/g, "")
-        .replace(/<core-memory-refresh>[\s\S]*?<\/core-memory-refresh>\s*/g, "")
-        .replace(/<system>[\s\S]*?<\/system>\s*/g, "")
-        .trim(),
-    )
+    .map((t) => {
+      let s = t;
+      // Injected context from memory system
+      s = s.replace(/<relevant-memories>[\s\S]*?<\/relevant-memories>\s*/g, "");
+      s = s.replace(/<core-memory-refresh>[\s\S]*?<\/core-memory-refresh>\s*/g, "");
+      s = s.replace(/<system>[\s\S]*?<\/system>\s*/g, "");
+      // Media attachment preamble (appears before Telegram wrapper)
+      s = s.replace(/^\[media attached:[^\]]*\]\s*(?:To send an image[^\n]*\n?)*/i, "");
+      // System exec output blocks (may appear before Telegram wrapper)
+      s = s.replace(/^(?:System:\s*\[[^\]]*\][^\n]*\n?)+/gi, "");
+      // Telegram wrapper — may now be at start after previous strips
+      s = s.replace(/^\s*\[Telegram\s[^\]]+\]\s*/i, "");
+      // "[message_id: NNN]" suffix
+      s = s.replace(/\n?\[message_id:\s*\d+\]\s*$/i, "");
+      return s.trim();
+    })
     .filter((t) => t.length >= 10);
 }

extensions/memory-neo4j/index.ts

@@ -1184,6 +1184,20 @@ const NOISE_PATTERNS = [
   /^[\p{Emoji}\s]+$/u,
   // System/XML markup
   /^<[a-z-]+>[\s\S]*<\/[a-z-]+>$/i,
+
+  // --- System infrastructure messages (never user-generated) ---
+  // Heartbeat prompts
+  /Read HEARTBEAT\.md if it exists/i,
+  // Pre-compaction flush prompts
+  /^Pre-compaction memory flush/i,
+  // System timestamp messages (cron outputs, reminders, exec reports)
+  /^System:\s*\[/i,
+  // Cron job wrappers
+  /^\[cron:[0-9a-f-]+/i,
+  // Gateway restart JSON payloads
+  /^GatewayRestart:\s*\{/i,
+  // Background task completion reports
+  /^\[\w{3}\s+\d{4}-\d{2}-\d{2}\s.*\]\s*A background task/i,
 ];
 
 /** Maximum message length — code dumps, logs, etc. are not memories. */