metadata, and filtering on it

Identifiers say whose memory it is. Metadata says what kind of memory it is, and it is the difference between a search that finds the right thing and one that finds four things.

from pretend_mem0 import memory

shop = memory()
shop.add("Deliver to my office.", user_id="ravi",
         metadata={"topic": "delivery"})
stored = shop.get_all(filters={"user_id": "ravi"})["results"][0]
print(stored["memory"], "|", stored["metadata"])

Any dictionary you like, stored beside the memory and given back with it. Mem0 does not interpret it; it is yours.

Filtering on it

from pretend_mem0 import memory

shop = memory()
shop.add("Deliver to my office.", user_id="ravi", metadata={"topic": "delivery"})
shop.add("I prefer email updates.", user_id="ravi", metadata={"topic": "contact"})

found = shop.search("office", filters={"user_id": "ravi", "topic": "delivery"})
for hit in found["results"]:
    print(hit["memory"])

The filter and the question work together: the filter narrows the set, then the question ranks what is left. One without the other is weaker than both.

What to put in it

• A topic or category, so an assistant handling a delivery question does not have to read preferences about anything else.
• Where the fact came from: a chat, a form, an import. You will want this the first time a memory is wrong.
• How confident you are, if some of your memories are guesses and others are settings the user typed.
• Anything you would otherwise have to parse back out of the sentence.

It does not change the text

Metadata is not embedded. Tagging a memory topic: delivery does not make it match the word delivery in a search; only the memory text is embedded. The tag filters, the text ranks, and confusing the two produces searches that quietly return nothing.