Bug Description

After upgrading to claude-mem v8.5.2, the plugin created 13,000+ orphaned observer session files over 2 days (Dec 31 - Jan 2). Each file contains only the initialization message "Hello memory agent, you are continuing to observe..." with no actual observations.

This is likely a severe side effect of the startup-recovery mechanism reported in #510.

Impact

• Token consumption: Each observer session calls the LLM API even though no actual work is being observed
• Financial cost: Thousands of unnecessary API calls
• Storage waste: 13,000+ .jsonl files created
• UX degradation: claude -r command shows thousands of useless sessions
• Database integrity: Database shows normal session counts (~30/day) but filesystem has 13,000+ files

Statistics

File System Analysis (.jsonl session files)

Database Analysis (sdk_sessions table)

Key Finding: Database shows normal activity (~13-30 sessions/day), but filesystem has 13,000+ files. This indicates observer sessions are being created but not properly tracked in the database.

Session File Content

Each orphaned session file contains only:

{
  "type": "queue-operation",
  "operation": "dequeue",
  "timestamp": "2026-01-01T16:30:42.480Z",
  "sessionId": "[session-uuid]"
}

{
  "parentUuid": null,
  "type": "user",
  "message": {
    "role": "user",
    "content": "Hello memory agent, you are continuing to observe the primary Claude session.\n\n<observed_from_primary_session>..."
  },
  "sessionId": "[session-uuid]"
}

No actual tool executions or observations are recorded.

Environment

• claude-mem version: 8.5.2
• Claude Code version: 2.0.76
• OS: macOS (Darwin 25.2.0)
• Provider: Custom API endpoint (via ANTHROPIC_BASE_URL)
• Worker service: Running on port 37777
• Start date of issue: December 31, 2025

Root Cause Hypothesis

Based on Issue #510, the startup-recovery mechanism in worker-service.ts:748 always uses sdkAgent regardless of CLAUDE_MEM_PROVIDER setting:

session.generatorPromise = this.sdkAgent.startSession(session, this)

Hypothesis: When worker restarts:

1. startSessionProcessor() is called for each pending session
2. Each call creates a NEW observer session file
3. The observer session initializes with "Hello memory agent..." message
4. But if there's no actual activity to observe, the session remains empty
5. This creates a loop where every recovery attempt spawns more empty observer sessions

Logs Evidence

From claude-mem-2025-12-31.log:

[INFO ] [SYSTEM] Auto-recovered 4 sessions with pending work {totalPending=4, started=4, sessionIds=2833,2910,2916,2924}

Multiple recovery attempts may have created multiple observer sessions for the same pending work.

Steps to Reproduce

1. Install claude-mem v8.5.2
2. Use Claude Code normally for a few sessions
3. Restart the worker service (either manually or due to version mismatch)
4. Observe the creation of numerous .jsonl files in ~/.claude/projects/-[username]/
5. Check database: sqlite3 ~/.claude-mem/claude-mem.db "SELECT COUNT(*) FROM sdk_sessions WHERE status='active'"
6. Run claude -r to see hundreds/thousands of "memory agent" sessions

Expected Behavior

• Startup-recovery should:
  • Resume existing sessions, not create new ones
  • Respect the CLAUDE_MEM_PROVIDER setting
  • Clean up failed recovery attempts
• Orphaned session files should not be created
• Database and filesystem should be in sync

Actual Behavior

• Startup-recovery creates thousands of observer session files
• Files exist in filesystem but not properly tracked in database
• Each observer session makes an LLM API call with the initialization prompt
• Sessions appear in claude -r list, cluttering the UI

Related Issues

• Startup-recovery ignores CLAUDE_MEM_PROVIDER setting, always uses Claude SDK #510: Startup-recovery ignores CLAUDE_MEM_PROVIDER setting (likely root cause)
• 「Bug」During the execution of tasks using claude code, cmd pop-ups kept appearing and then closed after a while, and port 37777 became inaccessible during the process. #494: cmd pop-ups kept appearing, port 37777 became inaccessible (related symptom)

Temporary Workaround

Disabled auto-observation in ~/.claude-mem/settings.json:

{
  "CLAUDE_MEM_CONTEXT_OBSERVATIONS": "0",
  "CLAUDE_MEM_CONTEXT_FULL_COUNT": "0",
  "CLAUDE_MEM_CONTEXT_SESSION_COUNT": "0",
  "CLAUDE_MEM_DISABLED": "true"
}

Then killed all worker processes:

pkill -f "claude-mem.*worker-service"
pkill -f "claude-mem.*mcp-server"

Cleanup Commands

To remove orphaned session files (backup first!):

# Backup
mkdir -p ~/.claude/backup/$(date +%Y%m%d)
find ~/.claude/projects/-[username] -maxdepth 1 -name "*.jsonl" -newermt "2025-12-31" -exec mv {} ~/.claude/backup/$(date +%Y%m%d)/ \;

# Mark database sessions as completed
sqlite3 ~/.claude-mem/claude-mem.db "UPDATE sdk_sessions SET status='completed', completed_at=datetime('now') WHERE status='active';"

Additional Notes

• Issue started after plugin update on Dec 31, 2025
• Previous version (before v8.5.2) did not have this problem
• The problem persists even after stopping/starting the worker
• Token consumption is significant: each observer session makes at least one API call

Severity

HIGH - This bug:

• Wastes API tokens/money
• Degrades user experience significantly
• Creates storage waste
• May affect all users who upgraded to v8.5.2
• Makes the claude -r command unusable

Request

Please investigate the startup-recovery mechanism and fix the orphaned session creation bug. A hotfix would be greatly appreciated given the financial impact.