slm remember xxx wait for a long time but seems no response

[RanchoCooper]

slm remember xxx wait for a long time but seems no response

[varun369]

Hi @RanchoCooper — thanks for reporting this. This is fixed in v3.3.19 (just released).

What happened

On first use, slm remember / slm warmup triggers the embedding model download (~500MB). The worker subprocess ran with stderr suppressed, so you saw no download progress — just a hang until the 180s timeout.

What's fixed in v3.3.19

1. Interactive setup wizard — runs automatically on first use. Downloads both embedding model and reranker model with visible progress before you use any command.
2. Clear timeout message — if the worker ever times out, you now get actionable guidance (slm setup / slm doctor).
3. 90x recall speedup — recall went from ~60s to <1s (unrelated bottleneck we also found and fixed).

How to upgrade

# npm
npm install -g superlocalmemory@latest

# pip
pip install --upgrade superlocalmemory

After upgrading, the setup wizard will auto-run on your next slm command and pre-download all models. You should never see a hang again.

If you still experience issues after upgrading, please run slm doctor and share the output.

[RanchoCooper]

ok, after upsrading, it seems took about 15 seconds per store operation

slm remember "I am xxx"

it's a little slow 😞

[varun369]

Hi @RanchoCooper — we shipped v3.3.20 with more fixes on top of the previous update.

What's new in v3.3.20

• slm remember is now instant (0.3s) — processes in background by default
• slm setup wizard auto-runs on first use with visible model downloads
• 90x recall speedup — from ~60s to <1s
• Non-blocking MCP — remember/recall no longer freeze your IDE

Upgrade

npm install -g superlocalmemory@latest
# or
pip install --upgrade superlocalmemory

How SLM is designed to be used

The CLI (slm remember, slm recall) works for quick tests, but SLM really shines when connected as an MCP server to your AI coding assistant. Here's how it works:

1. Install SLM — npm install -g superlocalmemory (installs everything including Python deps)
2. Connect to your IDE — Run slm connect (auto-detects and configures Claude Code + others)
3. Just code — SLM hooks into your AI assistant's lifecycle:
   • Session start: Auto-injects relevant memories from past sessions
   • During session: Auto-observes decisions, bugs, preferences you discuss
   • Session end: Auto-saves a rich session summary
   • Every recall: 6-channel retrieval + cross-encoder reranking + Fisher-Rao similarity
4. It learns over time — After 200+ signals, adaptive ML reranking kicks in. Your memory gets smarter the more you use it.

Think of SLM as a long-term memory layer for your AI agent — not a CLI database. The CLI commands (slm remember, slm recall) exist for debugging and quick tests, but the MCP integration is where the magic happens.

MCP Setup

For Claude Code / Claude Desktop — automatic:

slm connect
slm hooks install

For other IDEs (Cursor, Windsurf, Antigravity, Gemini CLI, etc.) — add this to your MCP config:

{
  "superlocalmemory": {
    "command": "slm",
    "args": ["mcp"],
    "env": {}
  }
}

Config file locations:

• Cursor: .cursor/mcp.json in your project root
• Windsurf: .windsurf/mcp.json or ~/.codeium/windsurf/mcp_config.json
• Antigravity: ~/.gemini/mcp_config.json
• Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json

Once connected, the AI agent gets 17 memory tools (remember, recall, observe, session_init, etc.) and your memory builds automatically as you work.

slm dashboard    # Web UI at localhost:8765 to see all your memories

Let us know if you run into anything on v3.3.20!

[RanchoCooper]

hi，@varun369

UI not found

remember is async operation now
but recall still took 19 seconds

and another bug👆🏻

[RanchoCooper]

Additionally, I'm very interested in the SuperLocalMemory project and would like to contribute code. Is that possible?

[varun369]

Hi @RanchoCooper — thank you for the detailed feedback and screenshots! We've addressed all the issues you reported in v3.3.21 (just released).

What's fixed

1. Dashboard UI — The dashboard assets are now properly bundled in the package. After upgrading, slm dashboard should show the full UI.

2. slm remember reliability — We redesigned the async store pipeline. Your data is now written to a safe local store immediately (0.1s), then processed in the background. Even if the background processing fails, your memory is preserved and automatically retried on the next operation. No more silent data loss.

3. slm recall speed — We added a background daemon that keeps the engine warm. First call in a session may take a few seconds (one-time warmup), but all subsequent calls are sub-second (~600ms). The daemon starts automatically on your first CLI use.

4. SpreadingActivation error — Fixed. The error you saw ('NoneType' object has no attribute 'search') is resolved.

How to upgrade

npm install -g superlocalmemory@latest
# or
pip install --upgrade superlocalmemory

After upgrading, try:

slm remember "I am Chinese"
slm recall "Chinese"
slm dashboard

Contributing

We'd love to have you contribute! Please feel free to fork the repo and submit PRs — I'll be happy to review and merge them. Focus areas where contributions would be especially welcome: i18n improvements, CLI experience, dashboard enhancements, and documentation.

In a few weeks, once we wrap up the current paper we're working on, I'd also be happy to add you as a direct contributor to the repository.

Thank you for being an early user and for taking the time to report these issues — it really helps make SLM better for everyone.

[RanchoCooper]

recall still token a lot of time

[Tyrin451]

Hi @varun369,

I'm currently running Python 3.12 on Windows 11 with the 3.3.28 version of the superlocalmemory.

I'm experiencing the exact same issue reported by @RanchoCooper in this first post. It seems to be a compatibility problem on my end as well.

Do you have any updates or a workaround for this ?

[varun369]

Hi @Tyrin451 — thanks for reporting this. We're aware of the issue on Windows 11 + Python 3.12 and are actively investigating.

The original hang was caused by the embedding model download running silently without progress output — this was fixed on macOS in earlier versions, but Windows subprocess handling has different behavior that we need to address separately.

We'll update this thread shortly with a fix for Windows users.

In the meantime, if you haven't already, try running:

slm setup

This forces the interactive model download with visible progress. If that completes successfully, subsequent slm remember calls should work.

---

Also shipping v3.3.29 today with fixes for two broken MCP tools (forget dry_run + run_maintenance) — contributed by @M74X. Upgrade with:

npm install -g superlocalmemory@latest
# or
pip install --upgrade superlocalmemory

[varun369]

@Tyrin451 — one more thing: if you're up for it, we'd welcome a contribution here! Since we develop primarily on macOS, having someone with a Windows machine debug and fix the subprocess handling would be incredibly valuable.

The likely culprit is in the embedding worker spawning — Windows doesn't support SIGTERM and uses CreateProcess instead of fork, so our process management needs Windows-specific handling.

Key files to look at:

• src/superlocalmemory/commands.py (CLI commands including remember)
• src/superlocalmemory/infra/process_reaper.py (process management)
• src/superlocalmemory/worker/embedding_worker.py (the worker that gets spawned)

Fork the repo, reproduce the issue, and submit a PR — I'd be happy to review and merge. @M74X just contributed a great fix today (PR #13) that shipped in v3.3.29, so the contribution workflow is proven and smooth.