Why a dictionary is not memory

The obvious fix is a dictionary keyed by user. It works for about ten minutes, and the ways it breaks are exactly the features Mem0 has.

Here is that first attempt. Everything the customer says gets stored under his name.

NOTES = {}

def remember(user, said):
    NOTES.setdefault(user, []).append(said)

def recall(user, about):
    return [note for note in NOTES.get(user, []) if about in note]

remember("ravi", "I prefer email updates, not SMS.")
remember("ravi", "Thanks!")
print(recall("ravi", "email"))

It stored the useful sentence and it found it. It also stored Thanks!, which is not a memory, and it found the note only because the question happened to use the word email.

The first break: the question uses different words

print(recall("ravi", "contact"))
print(recall("ravi", "how should we reach him"))

Nothing. The fact is sitting right there and a substring search cannot reach it, because nobody says email when they ask how to contact someone. Matching text is not the same as matching meaning, and that gap is what an embedder closes.

The second break: facts change

Ravi changes his mind and asks for SMS after all. The dictionary now holds both, contradicting each other, and neither is marked as the current one. Whatever you search for, you get an answer that might be a year out of date.

The third break: it never stops growing

Every thanks, every ok, every greeting is in there forever. After a thousand conversations the useful facts are a tiny fraction of the pile, and every search has to wade through the rest.