openclaw - 💡(How to fix) Fix [Bug]: TUI startup creates in-process memory watcher that recursively scans / (entire root filesystem), freezing terminal for ~8 minutes [2 comments, 2 participants]

Bug type

Crash (process/app exits or hangs)

Beta release blocker

No

Summary

openclaw version: 2026.4.29 (commit a448042c) platform: macOS 26.4.1 (Darwin 25.4.0, arm64) node: v25.9.0 mode: local embedded (no gateway) install path: /opt/homebrew/lib/node_modules/openclaw

summary

openclaw chat (and tui --local) creates a long-lived in-process chokidar watcher on TUI startup that recursively walks / (the root filesystem) with three concurrent workers. the TUI is unresponsive — including to Ctrl-C — for the duration of the walk (~8–10 minutes on a 1TB disk). after the scan completes, the TUI works normally.

this contradicts the documented behavior in docs/concepts/memory-qmd.md, which says "Gateway startup does not initialize QMD by default, so cold boot avoids importing the memory runtime or creating the long-lived watcher before memory is first used." TUI startup appears to take a different path that does create the watcher, and with a misresolved root.

root cause (high confidence)

dist/qmd-manager-D9gNmXKa.js constructs the watcher path as:

  resolveCollectionWatchPath(colle ction) {
    return path.join(path.normalize(collect ion.path), collection.pattern);
  }

with default pattern "**/*.md" (per dist/memory-mxD4Jzcs.js).

if collection.path is empty, "/", or otherwise unresolved, the watcher target resolves to /**/*.md, which chokidar implements by recursive readdir+stat. the documented ignore list (.git, node_modules, etc.) doesn't help when the root is the entire disk — there are millions of files outside those directories on a real machine.

asks

1. fix the path resolution. collection.path should never resolve to / for the default workspace collection. add a guard that throws if collection.path is falsy, root, or homedir-without-subpath.
2. honor the documented behavior. TUI startup should not initialize the QMD watcher; defer until first memory use.
3. add a config kill-switch. something like memory.qmd.watcher.enabled: false so users with a triggered bug can opt out without disabling memory entirely.
4. fail loud, not silent. if the watcher detects it's about to scan / or $HOME, log a fatal warning and abort instead of starting the walk.

workspace details (for repro)

  ~/Gits/sams-claw/
  ├── AGENTS.md
  ├── HEARTBEAT.md
  ├── IDENTITY.md
  ├── MEMORY.md
  ├── SOUL.md
  ├── TOOLS.md
  ├── USER.md
  └── memory/
      └── 2026-05-01.md

config has no memory.* block — all defaults.

Steps to reproduce

reproduce

1. fresh openclaw 2026.4.29 install on macOS, default config
2. workspace at ~/Gits/sams-claw containing MEMORY.md + memory/2026-05-01.md
3. from any cwd: openclaw chat
4. observe: terminal freezes, no input accepted, no Ctrl-C
5. in another terminal: lsof -p <tui-pid> | grep DIR shows three / handles
6. wait ~8 minutes; TUI eventually becomes interactive

Expected behavior

expected

per the docs: TUI/gateway cold start should not create the long-lived in-process watcher. memory features should be lazily initialized when first used.

if the watcher is created at startup, its path must resolve to a workspace-relative scope (e.g., ~/Gits/sams-claw/memory/**/*.md plus ~/Gits/sams-claw/MEMORY.md), not /.

Actual behavior

TUI freezes for 8+ mins; see evidence below.

OpenClaw version

2026.4.29 (commit a448042c)

Operating system

macOS 26.4.1 (Darwin 25.4.0, arm64)

Install method

npm

Model

anthropic/claude-opus-4.7

Provider / routing chain

openclaw -> anthropic

Additional provider/model setup details

No response

Logs, screenshots, and evidence

### evidence

1. process state during freeze


  PID 47879  STAT R+  CPU 99.2%  ETIME 01:18  openclaw-tui


2. open file descriptors — three parallel scans of /


  $ lsof -p 47879 | awk '$5=="DIR" && $4 ~ /r$/'
  node 47879 sam 6r DIR 1,18 704 2 /
  node 47879 sam 8r DIR 1,18 704 2 /
  node 47879 sam 10r DIR 1,18 704 2 /


cwd was /tmp for this run — the scan root is not derived from cwd.

3. stack sample (10s, all 7946 samples on main thread)


  DispatchQueue_1: com.apple.main-thread (serial)
    start
    → node::Start
    → node::SpinEventLoopInternal
    → uv_run → uv__io_poll → uv__async_io
    → uv__work_done                                          ← libuv work queue
    → node::MakeLibuvRequestCallback<u v_fs_s, AfterStat>
    → node::fs::AfterStat                                    ← fs.stat callback
    → InternalCallbackScope::Close
    → MicrotaskQueue::PerformCheckpoin tInternal
    → Builtins_PromiseFulfillReactionJob
    → Builtins_AsyncFunctionAwaitResolveClosure
    → Builtins_ArrayMap                                      ← .map() over
readdir entries
    → string concatenation + StringCharCodeAt               ← path joining


every sample is in the same fs.stat → microtask → ArrayMap → string-build chain.
classic recursive readdir+stat pattern.

4. heap footprint


  Physical footprint: 974.1M (peak 974.4M)


~1GB of string objects, consistent with accumulating directory paths during a
deep walk.

Impact and severity

No response

Additional information

No response