infer=False: storing what was actually said

Everything so far has gone through the model. Sometimes you do not want a judgement, you want the sentence, and there is a flag for that.

from pretend_mem0 import memory

shop = memory()
result = shop.add("Ravi prefers email updates. Thanks!", user_id="ravi",
                  infer=False)
for stored in result["results"]:
    print(stored["memory"])

The whole string, including the thanks, stored exactly as given. No model was called, nothing was split, nothing was dropped.

When that is what you want

When you already know the fact. A user ticks email me in a settings page. There is nothing to extract, and asking a model to think about it is slower, costlier and occasionally wrong.

When the text is already a summary. If another part of your system has written customer prefers email, delivery to office, putting that through an extractor only risks it coming out worse.

When you are testing. Storing exactly what you meant to store makes it obvious whether a search problem is in the search or in the extraction.

What you give up

from pretend_mem0 import memory

shop = memory()
shop.add("Ravi prefers email updates. Thanks!", user_id="ravi", infer=False)
for hit in shop.search("email", filters={"user_id": "ravi"})["results"]:
    print(round(hit["score"], 2), "|", hit["memory"])

The memory is one lump containing the useful part and the noise, so the score is lower than it would be for a clean fact, and everything you search for drags the thanks along with it. Over a few hundred conversations that difference decides whether search works.