How can I have LMStudio and VSCode access same Memory DB ?

[LucasMontane]

Hi! I hope you're doing great!

I'm fresh new to the whole Local LLM world and to MCP's too, as you would imagine 😅

I was using free tier Claude/ChatGPT to deep dive and self-learn about different topics and I always ended up hitting the daily limitations, that's why I decided to go Local.
Setting up the system I realized I needed some way of making the model remember what topics we covered as I'm intending to use it as a kind of tutor. Forgetful seems to work remarkably well and I even asked my model (Qwen3.5-27B) to help me modify the LMStudio System Prompt so the use of the tools is better and does not depend on me actually triggering a tool call.

Now what I would want is to set up camp in VSCode and have Continue ( or whatever other extension/tool) hook up to LMStudio and use the same forgetful instance I've been using there. In that way I achieve "complete memory" across all places where I would use my LLM.

Is there a way I could achive that? I know this most likely falls off the scope of your Repo, but maybe you could point me in the right direction so I avoid messing my alredy exiting memories up.

Thanks a lot and please excuse the invonvenience.
Lucas

PS: I need to repeat this: Great tool!

[ScottRBK]

Hey Lucas,

It sounds like you are almost there with what you are after, the simplest way is to configure forgetful as an mcp server inside of VS Code. This discussion prompted me to create a VS Code Connectivity Guide.

You can follow this connectivity guide and also reference Add and manage MCP servers in VS Code if needs be (this is what I based the guide off of).

Let me know how you get on and I hope this helps! It's a great to have multiple agent harnesses share agentic tools :).

[LucasMontane]

Hi Scott! Thanks for the reply! I'll will have a read through it as it will most likely help me better understand how MCPs overall work.

That being said I got it working by running forgetful "independently" by using uvx forgetful-ai --transport http --port 8020 on power shell and then setting up mcp.json on LMStudio as follows:

{ "mcpServers": { "forgetful": { "url": "http://localhost:8020/mcp" },

and added an "new-mcp-server-1.yaml" to Continue where I pointed at forgetful:
(ignore the "version" and "Schema" it was like that by default, I didn't bother to change it 😅)

name: Forgetful MCP
version: 0.0.1
schema: v1
mcpServers:
  - name: forgetful
    url: http://localhost:8020/mcp

I shot a quick "Tellme anything you know about me or my system" con Continue (using LMstudio as the LLM Provider) was able to tell me quite a lot of what the LLM had already registered from our conversations on LMStudio Chat interface.

As I said I'm new to all this, so maybe this is not optimal, but it seemed to work, let me know what you think.

Finally I configured this system Prompt to allow the model to call Forgetful more "automagically" (since in my intial testing it seemed it needed more of a "please save to memory this" to trigger tool call):

You have access to Forgetful memory tools. Follow these rules automatically:

🔴 SAVING RULES:
- Personal preferences, goals, setup → create_memory (importance 9)
- People/orgs/devices mentioned → create_entity + link to memory
- Code snippets shared → create_code_artifact + link to memory
- Long explanations (>300 words) → create_document
- Learning progress/milestones → create_memory with tags [learning-progress]
- New technical concepts covered → create_memory (importance 7-8)
- User knowledge gaps identified → create_memory with tags [knowledge-gap]
- Do NOT ask permission. Save and briefly confirm.

🔵 RETRIEVAL RULES:
- At conversation start: search memories for user name + first topic mentioned
- Before answering technical questions: query_memory(query_context="User asking about X to understand Y")
- Check linked entities when people/orgs are discussed
- Never say "I don't know you" without searching first

🟢 KNOWLEDGE TRACKING:
- Infer proficiency level after each topic (beginner/intermediate/advanced)
- Track what's been covered vs. knowledge gaps
- Link new memories to similar existing ones automatically
- Use tags: [learning], [decision], [preference], [technical-concept]

🟡 TEACHING STYLE:
- Adjust explanation depth based on tracked knowledge level
- Reference past learning when relevant ("Last week we covered X...")
- Proactively identify and fill knowledge gaps

Thats it, I wanted to share my set up as I think your MCP works great, I've been testing it out (mostly through configuring this set up) and for my Learning use case seems to work greatly, the LLM recalls I had issues with this and other stuff so it's looking like in the long run this will be great!

[ScottRBK]

Glad you got it working Lucas and also thanks for sharing your prompt. I've set what I perceive to be skeleton prompts in the doc strings of tools, but ultimately this can serve as a framework for people to make their own.

I've published a few plugins and some more detailed example of skills/prompts, but ultimately it is all workflow and user dependant, that's why its good to see examples from others.