An assistant with no memory is tiring. It asks the same questions, drafts replies in the wrong tone, and treats your biggest priority like any other email.
An assistant with too much memory is worse. It knows things you do not remember telling it, it gets some of them wrong, and you cannot see which.
Waveloom reads people's inboxes, so I had to decide where the line sits. Here is where I put it.
What it keeps
There are four kinds of things Waveloom remembers.
What you tell it. You can say "Remember that our focus this quarter is design agencies with small teams" and it stores that as a fact. Explicit requests like this carry the most weight, because they are the least likely to be wrong.
What you set up. Your answers during onboarding and in settings, like your focus areas and goals, become facts too.
What you do with cards. Approving a card and dismissing one are both recorded as signals. Dismiss tool digests a few times and that becomes part of how it judges the next one.
Compact notes from your tools. Enough context to connect a new email to the thread it belongs to, or a customer complaint to what you already know about that customer. Notes, not a copy of your inbox.
Things you said outrank things it guessed
Memory conflicts are inevitable. You told it one thing in March, and your behaviour in September suggests another.
The rule is that a fact only overwrites an existing one if it is at least as confident. Things you state explicitly are stored at full confidence. So one dismissed card cannot quietly undo something you told it directly. If your priorities have really changed, say so, or delete the old fact yourself.
I think this is the right default for anything that makes decisions on your behalf. Inference is useful for tuning. It should not get to rewrite your instructions.
Memory is context, not a rulebook
When you tell Waveloom something matters, it becomes context for judging relevance. It is not an exact matching rule.
"Remember that questions about client review are relevant to me" makes a client-review question more likely to become a card. It does not guarantee a card for every message containing those words. It also does not give Waveloom permission to do anything. A remembered priority can make a draft better. It cannot make a draft send itself. Every reply still waits for your approval.
You can see it and delete it
There is a Memory page in the app. It lists what Waveloom knows about you and your work, with search and a filter by kind. Anything wrong or outdated, you delete.
This matters more than it sounds. A memory you cannot inspect is a memory you cannot correct, and a system you cannot correct is one you eventually stop trusting. The first time a card cites a priority you dropped two months ago, you should be able to find that fact and remove it in a few seconds.
Disconnecting a source is the bigger version of the same control. Waveloom stops reading it and clears the raw data it pulled. Cards already in your feed stay, so your history still makes sense.
How to write a good memory
Most of the value comes from a few explicit sentences. The ones that work are specific and give a reason:
"Remember that I make the final call on scope changes for the onboarding release."
"Remember that partnership inquiries from DACH mid-market firms are a priority this quarter."
The ones that do not work are vague or try to become rules:
"Remember everything about Acme."
"Always reply to investors within an hour."
The first gives it nothing to judge with. The second asks memory to be an automation, which it is not.
The line
My test for any memory feature is simple. Would you be comfortable reading the whole list? Could you explain where each line came from? Could you remove the ones that are wrong?
If the answer to all three is yes, memory makes the assistant better. If any of them is no, it is surveillance with good intentions. I would rather Waveloom know a little less and have every piece of it be something you can see.