I spent the Saturday assembling a cedar planter, catching up with a college friend from IIT Roorkee, and exploring Anthropic's new model context protocol for their AI tool Claude. 

Yup, that's how I roll!

Claude now has richest set of MCP servers that can be integrated into Claude to augment its capabilities to, for example, search the web (think of it as a sort of RAG), search your filesystem, search your Google Drive, and best of all, save memories about your preferences that Claude can recall and leverage in future exchanges. Think of MCP servers as Anthropic's open-source framework aimed providing a bridge for AI agents or agentic AI to access real world data.

In large language models (LLMs), "context" refers to the information the model can "remember" during a single interaction — typically the prompt, any system instructions, and previous exchanges. This is often measured in tokens.

One more thing, unless you're running a quantum computer, it does take Claude several seconds to actually launch a new server before it shows up in the list of "MCP tools available" at the bottom left of the prompt window.

  • Filesystem Server

    If you've had a long exchange with Claude on a particular topic and want to save a summary of the exchange, this is a great option to do so while saving yourself a few steps. Just tell Claude to save a summary of the exchange and it will place file in your Desktop folder on Mac (in text, markdown, or PDF). I usually save it as a markdown (.md) file and then convert it to PDF after cleaning it up a bit using the md2pdf tool on my Mac.

    You can also use this server to search your filesystem and read/write files.

  • Google Drive

    This integration needs work. Per my current setup, Claude is able to see the folders, but not the contents inside those folders.

  • Brave Search

    Claude's knowledge, as it will readily admit ends around late 2024. So, if you're looking for some late breaking knowledge, Claude won't have it. But by integrating with Brave (a browser API), it can deliver something very similar to retrieval-augmented generation (RAG) and extend Claude's knowledge to bring it into the present. 

  • Memory Server

    Finally, this is what motivated me to try the whole MCP thing. So far Claude has always told me that it doesn't have the ability to save memories from our chats. So, every chat starts from scratch, which is obviously not ideal. This is where Memory Server could come to the rescue.

    But from my testing so far, that value proposition is yet to be fully realized.

    The geek in me wanted to know where Claude is storing the memories, hopefully not on the cloud. Turned out you can tell it where to store the memory.json file, but it won't honor your configuration. No worries. I'm not one to accept defeat so easily.

    Where does Claude store your memories?

    It seems like when Memory Server was running locally (i.e. in a NodeJS sandbox), it was storing memories here:

    /Users/lamba/.npm/_npx/15b07286cbcc3329/node_modules/@modelcontextprotocol/server-memory/dist/memory.json

    When I installed it globally (npm install -g <module name>), it started storing memories here:

    /usr/local/lib/node_modules/@modelcontextprotocol/server-memory/dist/memory.json

    One kink (probably a bug) is that if you start a net new session and ask Claude, "what do you know about me?" it will come up empty. If you then give it a new memory to save, it will suddenly remember everything.

    In case you're wondering, why not keep the same chat going forever? Claude will remember what's in the chat, right? Right, except that when I tried that Claude eventually told me that the chat had gotten too big and I needed to start a new chat.

    Since I didn't want Claude to lose these memories and I want to potentially share these memories across computers, I replaced Claude's master file with a symbolic link that points to the master file in my filesystem that is synced to cloud via Dropbox. (I do the same with my Sublime configurations etc, just in case I lose my laptop or something.)

    Once you've fed Claude aspects of your personal life, you can start asking more nuanced and personal questions like "What movie should I watch tonight?" or "Where should I live after retirement?" And Claude will remember things you've told it about yourself even after you have to reboot your computer because of an OS upgrade and start a net new chat!

  • Puppeteer Server

    If you've used a tool like Selenium to automate website testing, you can use this tool to automate testing using plain language. Although I think I crashed Claude while doing this.

  • Basic Memory

    Another interesting option. This is more for saving summaries of conversations as markdown files and being able to tell Claude to pick up from where you left off.

    There's a nuance here. I couldn't figure out how to move the location for where this server saves "notes" by default. (The lingo, aka prompts, are critical with GenAI. If you don't use the precise keywords, you will get blank stares, not unlike the human world.)

    So, after making the mistake of trying to create a symlink in my Dropbox folder to point to the basic-memory location, I realised that not only is that wrong because Dropbox doesn't support folder-based symlinks, but also because I want the SOT (source of truth) to be in Dropbox. So what you always want to do it to move the master file/folder to Dropbox, and create a symlink in the location where your app (in this case Claude's MCP server called basic-memory) expects to find it. 
When using Claude (with or without the various extensions), it's important to realize that you need to go into Claude > Settings > General > Claude Settings to provide some global prompts. These are guidelines that you want Claude to follow for all interactions. For example, this is where you can tell Claude where to save the basic memory notes. If you don't start your filepath with a "/" (on Mac) it will likely assume a relative path and save everything under the Claude install directory. 
With all of the above in place, I can ask Claude to, for example, create a "note" (basic-memory lingo) with movie recommendations (memory server knows my preferred genres) currently streaming (again, memory server knows my streaming subscriptions) and rated 7 or higher on IMDB. The end result is a nicely formatted markdown file sitting in my preferred folder location.

In closing, I think these are great strides for Claude and for users of Claude. I went ahead and upgraded from free to pro at $200/year. Mostly because I kept getting capped on exchanges. Totally worth it in terms of the value I am getting back. However, I am hoping that Claude will be more seamlessly integrated with the MCP servers and the adoption curve will continue to get smoother.

Comments

Popular posts from this blog

Milwaukee Appliance Hand Truck

My Experiments with the PICAXE 08M2+

Utility trailer buying guide