Best way to store non-visible workbook info in Mem0

[chengtie]

Hello,

I’m building an AI Agent embedded in Excel. Users chat with the Agent inside Excel, and the Agent can manipulate their workbook. I’m adding a memory layer using Mem0.

Right now, I can add the user–assistant conversation directly to memory, e.g.:

import MemoryClient, { Message, MemoryOptions } from 'mem0ai';

const client = new MemoryClient('your-api-key');

const messages: Message[] = [
    { role: "user", content: "I love Italian food" },
    { role: "assistant", content: "Italian food is great!" }
];

const options: MemoryOptions = {
    user_id: "alex",
    metadata: { category: "food_preferences" }
};

client.add(messages, options)
    .then(result => console.log(result))
    .catch(error => console.error(error));

I’d also like to store workbook-related facts in memory, ideally in natural language (e.g., “Workbook abc.xlsx contains 3 worksheets: jan, feb, mar”). Since this information isn’t shown to the user (it’s system-derived), I’m unsure of the best place to put it when calling Mem0:

• As content in a user message?
• In metadata? (though it’s not necessarily a simple key–value structure, and may be natural language)
• As a separate memory entry with a custom category (e.g., workbook_topology)?
• Another recommended pattern?

What’s the best practice in Mem0 for storing non-visible workbook context so it’s retrievable later alongside conversational memories?

Thanks!

[parshvadaftari]

Hey @chengtie really great to hear that your building with Mem0! Love the Excel AI Agent use case. You can add the details into metadata something like this:

metadata: {
          category: "workbook_context",
          workbook_name: "abc.xlsx"
      }

This way you use metadata to filter/distinguish workbook memories from conversational ones. Keep them under the same user_id for unified retrieval when needed

[chengtie]

Hello @parshvadaftari , thank you for your reply. How would you add, for instance, Workbook abc.xlsx contains 3 worksheets: jan, feb, mar to the metadata?

[parshvadaftari]

You can play around with the key value pair in metadata, for this it can be a workbook_description or another and better option would be to add it inside the user's memories. As Mem0 is designed to extract meaningful facts from NLP content, it will semantically index this information for better retrieval which later can be used using search and get functions.

On top of this you can still use metadata for filtering (e.g., by workbook_name or category) for better precision.