memory-neo4j: filter open proposals and cron noise from memory

Open proposals ("Want me to...?", "Should I...?") are dangerous in
long-term memory because other sessions interpret them as active
instructions and attempt to carry them out. This adds:
- Attention gate patterns for cron delivery outputs and assistant proposals
- Extractor scoring rules to rate proposals/action items as low importance
- Sleep-cycle Phase 7 to retroactively clean existing noise-pattern memories

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

extensions/memory-neo4j/attention-gate.ts

@@ -51,6 +51,22 @@ const NOISE_PATTERNS = [
   // --- Conversation metadata that survived stripping ---
   /^Conversation info\s*\(/i,
   /^\[Queued messages/i,
+
+  // --- Cron delivery outputs & scheduled reminders ---
+  // Scheduled reminder injection text (appears mid-message)
+  /A scheduled reminder has been triggered/i,
+  // Cron delivery instruction to agent (summarize for user)
+  /Summarize this naturally for the user/i,
+  // Relay instruction from cron announcements
+  /Please relay this reminder to the user/i,
+  // Subagent completion announcements (date-stamped)
+  /^\[.*\d{4}-\d{2}-\d{2}.*\]\s*A sub-?agent task/i,
+  // Formatted urgency/priority reports (email summaries, briefings)
+  /(\*\*)?🔴\s*(URGENT|Priority)/i,
+  // Subagent findings header
+  /^Findings:\s*$/im,
+  // "Stats:" lines from subagent completions
+  /^Stats:\s*runtime\s/im,
 ];
 
 /** Maximum message length — code dumps, logs, etc. are not memories. */
@@ -171,6 +187,22 @@ const ASSISTANT_NARRATION_PATTERNS = [
   /^All (good|set|done)[!.]/i,
   // "Here's what changed" / "Summary of changes" (session-specific)
   /^(here'?s\s+(what|the|a)\s+(changed?|summary|breakdown|recap))/i,
+
+  // --- Open proposals / action items (cause rogue actions when recalled) ---
+  // These are dangerous in memory: when auto-recalled, other sessions interpret
+  // them as active instructions and attempt to carry them out.
+  // "Want me to...?" / "Should I...?" / "Shall I...?" / "Would you like me to...?"
+  /want me to\s.+\?/i,
+  /should I\s.+\?/i,
+  /shall I\s.+\?/i,
+  /would you like me to\s.+\?/i,
+  // "Do you want me to...?"
+  /do you want me to\s.+\?/i,
+  // "Can I...?" / "May I...?" assistant proposals
+  /^(can|may) I\s.+\?/i,
+  // "Ready to...?" / "Proceed with...?"
+  /ready to\s.+\?/i,
+  /proceed with\s.+\?/i,
 ];
 
 export function passesAssistantAttentionGate(text: string): boolean {

extensions/memory-neo4j/extractor.ts

@@ -358,6 +358,8 @@ KEY RULES:
 - AI assistant self-narration ("Let me check...", "I'll now...", "Done! Here's what changed...") is ALWAYS 1-3
 - System prompts, formatting instructions, voice mode rules are ALWAYS 1-2
 - Technical debugging details ("the WebSocket failed because...") are 2-4 unless they encode a reusable lesson
+- Open proposals and unresolved action items ("Want me to fix it?", "Should I submit a PR?", "Would you like me to proceed?") are ALWAYS 1-2. These are dangerous in long-term memory because other sessions interpret them as active instructions.
+- Messages ending with questions directed at the user ("What do you think?", "How should I handle this?") are 1-3 unless they also contain substantial factual content worth remembering
 - Personal facts about the user or their family/contacts are 7-10
 - Business rules and operational procedures are 7-9
 - Preferences and opinions expressed by the user are 6-8

extensions/memory-neo4j/sleep-cycle.ts

@@ -657,6 +657,63 @@ export async function runSleepCycle(
     }
   }
 
+  // --------------------------------------------------------------------------
+  // Phase 7: Noise Pattern Cleanup
+  // Removes memories matching dangerous patterns that should never have been
+  // stored (open proposals, action items that trigger rogue sessions).
+  // --------------------------------------------------------------------------
+  if (!abortSignal?.aborted) {
+    logger.info("memory-neo4j: [sleep] Phase 7: Noise Pattern Cleanup");
+
+    try {
+      const noisePatterns = [
+        "(?i)want me to\\s.+\\?",
+        "(?i)should I\\s.+\\?",
+        "(?i)shall I\\s.+\\?",
+        "(?i)would you like me to\\s.+\\?",
+        "(?i)do you want me to\\s.+\\?",
+        "(?i)ready to\\s.+\\?",
+        "(?i)proceed with\\s.+\\?",
+      ];
+
+      let noiseRemoved = 0;
+      const noiseSession = (db as any).driver!.session();
+      try {
+        for (const pattern of noisePatterns) {
+          if (abortSignal?.aborted) {
+            break;
+          }
+
+          const agentFilter = agentId ? "AND m.agentId = $agentId" : "";
+          const result = await noiseSession.run(
+            `MATCH (m:Memory)
+             WHERE m.text =~ $pattern
+               AND coalesce(m.userPinned, false) = false
+               AND m.category <> 'core'
+               ${agentFilter}
+             WITH m LIMIT 100
+             DETACH DELETE m
+             RETURN count(*) AS removed`,
+            { pattern: `.*${pattern}.*`, agentId },
+          );
+          noiseRemoved += (result.records[0]?.get("removed") as number) ?? 0;
+        }
+      } finally {
+        await noiseSession.close();
+      }
+
+      if (noiseRemoved > 0) {
+        onProgress?.("cleanup", `Removed ${noiseRemoved} noise-pattern memories`);
+      }
+
+      logger.info(
+        `memory-neo4j: [sleep] Phase 7 complete — ${noiseRemoved} noise memories removed`,
+      );
+    } catch (err) {
+      logger.warn(`memory-neo4j: [sleep] Phase 7 error: ${String(err)}`);
+    }
+  }
+
   result.durationMs = Date.now() - startTime;
   result.aborted = abortSignal?.aborted ?? false;