[Windows] No observations saved for 5 weeks; chroma-mcp CMD window can't be closed + port 37777 zombie

[hugor11]

Hi — I'm a heavy claude-mem user on Windows and I'm out of ideas. I've spent ~1 week and ~7 reinstall/fix attempts and the tool still doesn't persist anything. Posting here as a last resort because I've lost ~5 weeks of memory and I don't know what else to try.

Setup

• OS: Windows 11 Pro 10.0.26200 (x64)
• Node: v22.20.0 — Bun: installed — uv/uvx: installed
• Versions tried: 12.3.9, 12.4.3, 12.4.7, 12.4.8, 12.5.0 (current)
• IDE: Claude Code CLI (git-bash + PowerShell)

Three coupled problems

1. No observations persisted since 30-mar-2026 (today is 3-may).

• DB last meaningful write: 30-mar. WAL is touched daily but no rows added.
• Hooks ARE firing (logs daily in ~/.claude-mem/logs/), but they log "Worker failed to start before hook, handler will proceed gracefully" and then the data is silently dropped.
• Only 3 of the last 30 days logged any archive/observation event.
• ~/.claude-mem/transcript-watch.json only contains a watch for Codex (~/.codex/sessions/**/*.jsonl). There is no watch for Claude Code transcripts (~/.claude/projects/**/*.jsonl). I assume the installer should add this — but reinstalling never re-adds it (see import-history command fails with TypeError: path argument must be string #3).

2. chroma-mcp opens a persistent visible CMD window I cannot close.
Worker spawns it via:

cmd.exe /c uvx --python 3.13 chroma-mcp==0.2.6 --client-type persistent --data-dir C:/Users/hugor/.claude-mem/chroma

The CMD window stays open for the entire session. If I close it, chroma dies and logs flood with [CHROMA_SYNC] Batch add failed, chroma-mcp connection in backoff (249 errors in last 30 days).

3. Port 37777 zombies between Claude Code restarts.

• 17× [ERROR] Failed to start server. Is port 37777 in use? in last 30 days
• Worker falls back to 37778
• Only a full system reboot clears it
• I suspect this is caused by problem ACCIDENTALLY TAGGED #2: the orphan chroma CMD prevents clean shutdown, so the listening socket leaks.

Reinstall via npx claude-mem install --ide claude-code aborts

gyp ERR! find VS could not use PowerShell to find Visual Studio 2017 or newer
error: install script from "tree-sitter-kotlin" exited with 1

I don't have Visual Studio C++ build tools (and don't use Kotlin). The installer aborts on this native dep before any post-install step runs, so transcript-watch.json never gets regenerated. This is why ~7 reinstall attempts changed nothing on the config side.

What I've already tried

• Full uninstall + reinstall multiple times (each one aborts on tree-sitter-kotlin)
• Plugin upgrades across all 12.3.9 → 12.5.0
• Renamed hooks.json → hooks.json.disabled (mitigates bash-hook flashes; doesn't fix the chroma CMD or persistence)
• Switched to memsearch on 23-apr while waiting; came back on 26-apr
• Multiple system reboots
• Manual worker restarts via npx claude-mem restart

Questions

1. When --ide claude-code is used, is transcript-watch.json supposed to contain a claude-code watch entry? Could you share the expected schema so I can patch it manually while waiting?
2. Is there a way to make tree-sitter-kotlin (and other heavy native parsers) optional, or a --skip-native-parsers flag for the installer? On Windows without VS C++ build tools the install just dies.
3. Can chroma-mcp be spawned hidden / detached on Windows so the CMD window doesn't lock the user out? subprocess.CREATE_NO_WINDOW or start /b style.
4. Any recommended workaround so port 37777 releases cleanly between Claude Code restarts on Windows?

I have full logs (~/.claude-mem/logs/), transcript-watch.json, settings.json hooks section, and a DB dump available — happy to share whatever helps. I really want this working again. Thank you for any pointer 🙏