I spent weeks making my AI companion remember. The real feature was letting her forget.

I thought the hardest memory problem was recall.

How do I make an AI companion remember the important details?

Your name. Your routines. The things you mentioned three weeks ago. The small preferences that make a conversation feel less like starting over every time.

I spent a lot of time thinking about that.

Then I realized I was solving only half the problem.

If an AI companion can remember something about you, you should also be able to:

- see exactly what she remembers

- see where that memory came from

- correct it when she got it wrong

- delete it when you no longer want it there

So I added a Memory Book to Local Waifu.

Every memory is visible as an actual entry instead of being hidden inside an opaque “AI memory” system.

You can expand a memory and see the source messages behind it.

The UI literally says:

“Where she learned this”

That detail matters.

An AI can make a plausible inference that is still wrong. It can turn a temporary situation into a permanent fact. It can misunderstand sarcasm. It can remember something that was true six months ago but is not true anymore.

If you cannot inspect the source, you cannot really trust the memory.

I also added editing.

When you edit a memory, the text is updated and its embedding is regenerated. Otherwise, the database would contain one sentence for the user and a different semantic representation for the model.

And then there is the less exciting feature:

Delete.

Not “hide this for now.”

Not “ask support to remove it.”

Delete the memory entry after confirmation.

The most important part of this feature is also the least glamorous: the delete button runs against the memory record itself. It is not a fake UI promise.

There is one distinction I want to be very clear about:

Deleting a memory does not retroactively delete the original chat messages that created it.

The conversation and the extracted memory are separate data.

That separation is important because “forget this fact” and “delete our entire conversation” are not always the same request. A user should be able to understand that difference instead of trusting vague language like “your data is private.”

This changed how I think about AI companions.

The interesting problem is not only whether the model sounds warm, remembers your favorite things, or uses the right nickname.

The deeper question is:

Who owns the relationship history?

If the answer is “the company’s database,” then the user is borrowing the relationship.

If the answer is “the user can inspect, correct, export, and remove it,” then the product starts to feel more like something they actually control.

That is one reason I built Local Waifu as a desktop application with local data.

I do not want the companion to be a black box that quietly accumulates a profile of someone while pretending to be intimate.

I want the user to be able to open the memory layer and say:

“That is right.”

“That is wrong.”

“That was true once, but not anymore.”

“Forget this.”

The delete button is not the feature that looks best in a demo.

It does not produce a more impressive screenshot.

It does not make the model benchmark better.

But for a product that remembers personal things, I think it may be one of the most important features you can ship.

A companion should remember enough to feel continuous.

But the user should always have the final say in what remains.

I’m building Local Waifu around that idea:

An AI companion that lives on your computer, with a memory you can actually inspect and control.

https://localwaifu.com/