Custom Instructions - Mem0

Custom Instructions for Mem0

Custom instructions let you decide exactly which facts Mem0 records from a conversation. Define a focused prompt, give a few examples, and Mem0 will add only the memories that match your use case.

You’ll use this when…

• A project needs domain-specific facts (order numbers, customer info) without storing casual chatter.
• You already have a clear schema for memories and want the LLM to follow it.
• You must prevent irrelevant details from entering long-term storage.

Prompts that are too broad cause unrelated facts to slip through. Keep instructions tight and test them with real transcripts.

The custom_fact_extraction_prompt parameter has been renamed to custom_instructions. If you are upgrading from an older version, update your configuration accordingly.

Feature Anatomy

• Prompt instructions: Describe which entities or phrases to keep. Specific guidance keeps the extractor focused.
• Few-shot examples: Show positive and negative cases so the model copies the right format.
• Structured output: Responses return JSON with a facts array that Mem0 converts into individual memories.
• LLM configuration: custom_instructions (Python) or customInstructions (TypeScript) lives alongside your model settings.

Prompt Blueprint

1. State the allowed fact types.
2. Include short examples that mirror production messages.
3. Show both empty ([]) and populated outputs.
4. Remind the model to return JSON with a facts key only.

Configure It

Write the Custom Prompt

Python

custom_instructions = """
Please only extract entities containing customer support information, order details, and user information.
Here are some few shot examples:

Input: Hi.
Output: {"facts" : []}

Input: The weather is nice today.
Output: {"facts" : []}

Input: My order #12345 hasn't arrived yet.
Output: {"facts" : ["Order #12345 not received"]}

Input: I'm John Doe, and I'd like to return the shoes I bought last week.
Output: {"facts" : ["Customer name: John Doe", "Wants to return shoes", "Purchase made last week"]}

Input: I ordered a red shirt, size medium, but received a blue one instead.
Output: {"facts" : ["Ordered red shirt, size medium", "Received blue shirt instead"]}

Return the facts and customer information in a json format as shown above.
"""

TypeScript

const customInstructions = `
Please only extract entities containing customer support information, order details, and user information.
Here are some few shot examples:

Input: Hi.
Output: {"facts" : []}

Input: The weather is nice today.
Output: {"facts" : []}

Input: My order #12345 hasn't arrived yet.
Output: {"facts" : ["Order #12345 not received"]}

Input: I am John Doe, and I would like to return the shoes I bought last week.
Output: {"facts" : ["Customer name: John Doe", "Wants to return shoes", "Purchase made last week"]}

Return the facts and customer information in a json format as shown above.
`;

Keep example pairs short and mirror the capitalization, punctuation, and tone you see in real user messages.

See It in Action

Example: Order Support Memory

Python

m.add("Yesterday, I ordered a laptop, the order id is 12345", user_id="alice")

TypeScript

await memory.add("Yesterday, I ordered a laptop, the order id is 12345", { userId: "user123" });

Output

{
  "results": [
    {"memory": "Ordered a laptop", "event": "ADD"},
    {"memory": "Order ID: 12345", "event": "ADD"},
    {"memory": "Order placed yesterday", "event": "ADD"}
  ]
}

The output contains only the facts described in your prompt, each stored as a separate memory entry.

Example: Irrelevant Message Filtered Out

Python

m.add("I like going to hikes", user_id="alice")

TypeScript

await memory.add("I like going to hikes", { userId: "user123" });

Output

{
  "results": []
}

Empty results show the prompt successfully ignored content outside your target domain.

Best Practices

1. Be precise: Call out the exact categories or fields you want to capture.
2. Show negative cases: Include examples that should produce [] so the model learns to skip them.
3. Keep JSON strict: Avoid extra keys; only return facts to simplify downstream parsing.
4. Version prompts: Track prompt changes with a version number so you can roll back quickly.
5. Review outputs regularly: Spot-check stored memories to catch drift early.