Hi Clawdbot team,                                                                                                      
                                                                                                                        
 On macOS (snizServer), the clawdbot memory CLI subcommands appear to complete their work but then never exit:          
                                                                                                                        
 - clawdbot memory status --json prints valid JSON quickly, but the process stays running indefinitely.                 
 - clawdbot memory index prints “Memory index updated.” but doesn’t exit.                                               
 - clawdbot memory search --json "<query>" produces no output and hangs (20–40s+).                                      
                                                                                                                        
 This doesn’t look like a “dirty flag” semantics issue; it’s process lifecycle / open-handle behavior.                  
                                                                                                                        
 What we observed while debugging:                                                                                      
 - No lingering network connections: lsof -p <pid> -a -i shows no TCP/UDP sockets.                                      
 - The stuck process shows multiple KQUEUE file descriptors via lsof -p <pid>.                                          
 - sample <pid> shows the main thread sitting in uv__io_poll → kevent, which is the classic signature of a persistent   
 filesystem watcher (kqueue/fsevents/chokidar/FSWatcher) keeping the Node event loop alive.                             
                                                                                                                        
 We tested disabling memory watcher via config (agents.defaults.memorySearch.sync.watch=false):                         
 - It made memory status --json exit cleanly in one run,                                                                
 - but memory index and memory search still timed out/hung, so there may be additional watch/persistent handles started 
 in those paths.                                                                                                        
                                                                                                                        
 Hypothesis / likely root cause:                                                                                        
 - A filesystem watcher is started during memory manager initialization or indexing/search flows and is not being       
 closed (or is created as persistent), so one-shot CLI commands never terminate.                                        
                                                                                                                        
 Suggested fixes:                                                                                                       
 1. Don’t start watchers/interval timers in “one-shot CLI” contexts (only in long-running gateway/daemon mode).         
 2. Ensure any watcher/timer is non-blocking for CLI:                                                                   
     - chokidar: persistent: false and/or reliably await watcher.close()                                                
     - FSWatcher: .unref() or persistent: false                                                                         
 3. As a last resort: explicitly process.exit(0) after printing final output for memory status / CLI completion.        
                                                                                                                        
 Happy to provide additional artifacts if helpful (stdout, /tmp/clawdbot/clawdbot-YYYY-MM-DD.log, and a sample trace),  
 but we’re avoiding sharing vault content/snippets for privacy.                                                         
                                                                                                                        
 Thanks!                                                                                                                
                                                                                                                        
 — Brian