lmstudio

Mnemo Cortex — LM Studio Integration

Give your local LM Studio model persistent memory across chats and sessions. One config file, restart, done.

LM Studio added native MCP client support in v0.3.17. With this integration, your LM Studio chats — running entirely on your machine, on any tool-capable open-weights model — gain persistent semantic memory, brain-file file access, and (optional) cross-agent search alongside other agents on your network.

Prerequisites

• LM Studio v0.3.17 or later — earlier versions have no MCP client
• Node.js 18+ on your PATH (for the bridge)
• A running Mnemo Cortex server — see the main install guide. It can be on this machine (http://localhost:50001) or anywhere reachable on your network.
• A tool-capable model loaded in LM Studio. Tested working: Qwen3 (any size), Llama 3.2 3B Instruct, Mistral 7B Instruct v0.3, Hermes-3 Llama 3.1 8B. Models without tool-calling ability will appear to "succeed" while never actually calling Mnemo — see Gotchas below.

Install

1. Clone the bridge

git clone https://github.com/GuyMannDude/mnemo-cortex.git
cd mnemo-cortex/integrations/mcp-bridge
npm install

The bridge lives in mcp-bridge/ — it's the same Node service used by every Mnemo MCP integration (LM Studio, Claude Desktop, OpenClaw, etc.).

2. Edit mcp.json

Open LM Studio's MCP config file:

Platform	Path
Windows	%USERPROFILE%\.lmstudio\mcp.json
macOS	~/.lmstudio/mcp.json
Linux	~/.lmstudio/mcp.json

If the file doesn't exist, create it. Add this entry under mcpServers:

{
  "mcpServers": {
    "mnemo-cortex": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/mnemo-cortex/integrations/mcp-bridge/server.js"],
      "env": {
        "MNEMO_URL": "http://localhost:50001",
        "MNEMO_AGENT_ID": "lmstudio",
        "MNEMO_SHARE": "separate"
      }
    }
  }
}

Replace /ABSOLUTE/PATH/TO with where you cloned the repo. Adjust MNEMO_URL if your Mnemo server is remote.

3. Restart LM Studio

Fully quit LM Studio (not just close the window) and reopen. The MCP config is read at startup only.

Verify

1. Open a chat with a tool-capable model.
2. Click the MCP tab in the chat panel (right side of the chat input).
3. You should see mnemo-cortex listed with 9 tools:
   • 4 memory tools: mnemo_recall, mnemo_search, mnemo_save, mnemo_share
   • 5 Passport tools: passport_get_user_context, passport_observe_behavior, passport_list_pending_observations, passport_promote_observation, passport_forget_or_override

If BRAIN_DIR and/or WIKI_DIR are also set (see Optional), you'll see additional tools.

Quick functional test

In a new chat:

You: Save a note that I prefer concise replies.

The model should call mnemo_save (you'll see the tool invocation in the chat). Then start a new chat:

You: What do you remember about my preferences?

The model should call mnemo_recall and surface "concise replies." That round-trip across separate chats confirms persistence.

Optional: brain file and wiki tools

The bridge can expose project-context tools when you set extra env vars:

"env": {
  "MNEMO_URL": "http://localhost:50001",
  "MNEMO_AGENT_ID": "lmstudio",
  "BRAIN_DIR": "/path/to/your/mnemo-plan/brain",
  "WIKI_DIR": "/path/to/your/wiki"
}

When BRAIN_DIR points at an existing directory, these tools auto-register:

• read_brain_file, write_brain_file, list_brain_files
• opie_startup (session-start context bundle)
• session_end (writeback + brain commit)

When WIKI_DIR points at an existing directory, these tools register:

• wiki_search, wiki_read, wiki_index

Both are optional. Memory tools work without either. See the mnemo-plan template for a starter brain repo.

Gotchas

These are real, verified failure modes — not theoretical:

1. Non-tool-capable models silently fake success

This is the single biggest pitfall on LM Studio. Some open-weights models will narrate tool calls in their text response without actually invoking them:

Llama 3.1 8B (NOT tool-capable): "I've saved that to memory with id e4d3c9f1."

The memory ID is hallucinated. Nothing was saved. The model is performing what it thinks a tool call looks like in its training data, not actually emitting structured tool-use tokens that LM Studio's MCP client can parse.

Fix: Use a model with native tool-calling support. Qwen3 (any size), Llama 3.2 Instruct, Mistral 7B v0.3, and Hermes-3 are confirmed working.

Verify it's actually calling tools by opening the Tool Calls panel in LM Studio — real invocations show up there with structured arguments. If the chat says "saved" but Tool Calls is empty, the model faked it.

2. Bridge changes need a full LM Studio restart

LM Studio reads mcp.json only at app launch. If you edit the config (change MNEMO_AGENT_ID, point at a different server, etc.), close LM Studio fully — including from the system tray on Windows — and reopen. Reloading the model is not enough.

3. Server unreachable produces a mid-chat tool error

If Mnemo Cortex goes down while LM Studio is running, the next memory tool invocation will return an error to the model. Most tool-capable models handle this gracefully ("I can't reach memory right now, but..."). Some weaker models will get confused. Best practice: run a quick curl http://localhost:50001/health before opening LM Studio if you suspect server issues.

Sharing & Privacy

By default, each agent sees only its own memories. Cross-agent search is off.

Mode	MNEMO_SHARE=	Behavior
Separate (default)	separate or unset	Search restricted to own agent. mnemo_share toggles per-session.
Always	always	Cross-agent search always on. For trusted teams.
Never	never	Cross-agent search permanently off. Toggle blocked.

Use separate if you also run other agents (Claude Desktop, OpenClaw, Claude Code) and want LM Studio kept independent. Use always if you want LM Studio to read what those other agents have learned.

Environment Variables

Variable	Default	Description
MNEMO_URL	http://localhost:50001	Mnemo Cortex API address
MNEMO_AGENT_ID	openclaw (rename to lmstudio)	This agent's identity in the memory system
MNEMO_SHARE	separate	Cross-agent sharing mode
BRAIN_DIR	unset	Optional — enables brain-file tools when set
WIKI_DIR	unset	Optional — enables wiki tools when set

How It Works

LM Studio spawns the Mnemo Cortex bridge (mcp-bridge/server.js) as a child process using MCP stdio transport. When your LM Studio model invokes a memory tool, the bridge calls Mnemo Cortex's REST API:

• mnemo_recall → POST /context (your agent only)
• mnemo_search → POST /context (cross-agent gated by share mode)
• mnemo_save → POST /writeback
• mnemo_share → toggles session share state (no API call)

All requests have a 10-second timeout. The bridge itself logs to stderr (visible in LM Studio's developer console).

Workflow

For day-to-day use patterns, see the Session Guide. It covers when to recall, when to save, how to structure a brain file, and per-platform boot snippets.

Next Step

Read THE-LANE-PROTOCOL.md to learn the session ritual that makes Mnemo actually work.

License

MIT

---

Part of Mnemo Cortex by Project Sparks