Memory and Activity: what the assistant remembers

Besides what you write for it, the assistant remembers things on its own from your conversations. Remembered facts appear in the Memory tab, and what it worked on appears in Activity.

The remembering-on-its-own part has a problem worth knowing before rather than after.

Facts appear with no notice

After a few conversations, Memory goes from „no facts remembered” to a number. Nothing tells you something was written there.

Each fact carries its source and the quote it is based on. Both are worth reading, because the source can be the assistant itself.

The loop you need to know about

In a real test, the two facts remembered after four conversations had as their source sentences written by the assistant, not by the person. One came from a draft it had written, the other from one of its own chat messages.

It showed concretely in that test: the first remembered fact contained exactly the wrong address the person had just banned in the instructions. The instruction said one thing, memory had fixed another, and nothing on screen said which wins.

What to do in practice

1. Read Memory after every working session with the assistant, like a journal of what it thought it learned. It is the only defence against the loop
2. Delete what is wrong or redundant, from the bin on the right of each fact. A fact repeating a rule from your instructions is redundant: change the rule in one place and the other stays old
3. Fill Memory manually, with the add button. Facts you add stay until you delete them. The field itself states the right form: one sentence, in the present tense, about the brand or about your preferences

Memory is for preferences that do not fit in the instructions. Facts about your business belong in Knowledge.

What it does not keep

The assistant keeps things about your brand and about how your emails look. It keeps nothing about your subscribers as people.