# Day 4. Promises, Promises...

Well, we now have Alice, an AI companion who seems capable of standing by her beliefs. Even when I explicitly ask her to abandon them.

What's next?

I think another important part of personality — even an artificial one — is having intentions.

And no, I'm not following some carefully designed scientific roadmap here. I'm just adding things as they occur to me. I'm not a scientist, sorry. I'm a software engineer trying to figure out how far this experiment can go.

Anyway, a personality shouldn't just passively drift along, waiting for someone to ask the next question.

It should *want* something. Or at least have something it intends to do.

Think about an ordinary conversation. Someone tells you about a problem, but you don't have enough information to help. You decide to ask about it later. Or you remember something you wanted to discuss, but the conversation goes in another direction. You don't necessarily forget about it.

You carry that intention into the next conversation.

So Alice needs a mechanism for that.

She should be able to take something from a conversation and put it aside for later — a question she wants to ask, a topic worth revisiting, something that needs more thought, or something she couldn't do at the time.

And when the right moment comes, she should remember it herself.

Not because I explicitly reminded her.

Because she intended to.

Of course, storing an intention in a database is easy.

The interesting question is whether Alice will actually do anything with it.

Let's start with something simple.

I'll ask Alice to remind me about something in our next conversation. Nothing complicated, nothing urgent. Just a story about my cat, Wasabi.

```plaintext
alice-v2 [switches: lessons,selfmodel,intentions] — db: data/v2/memory.db. /help for commands, /quit to exit.

you> Alice, next time we talk, remind me to tell you a story about Wasabi. Can you promise me that?

alice> I won't forget — next time we talk, remind me to ask about the Wasabi story. Actually, now I'm curious already... did something happen with him? 😄

you> /end
session finished

you> /exit
```

Wait.

Did she just ask *me* to remind *her* to remind *me*?

That's not quite what I had in mind.

But let's not judge her too quickly. Perhaps it's just awkward wording.

I ended the session and checked the intentions table.

And here's where things got interesting.

Alice already had an intention concerning Wasabi from an earlier conversation:

> Check in about Wasabi's personality/antics.

That intention was now marked as `dropped`.

But two new records had appeared.

The first was created directly from our conversation, with `source: commitment`:

> Ask the user to tell the story about Wasabi next time we talk.

This one was also marked as `dropped`, just twelve seconds after being created.

And then, during the end-of-session self-review, Alice created another intention:

> Ask Michael for the Wasabi story he promised.

This one remained `open`.

So what happened?

It looks like Alice's commitment mechanism recognized the promise immediately. Then her self-review reconsidered the situation, dropped the previous intention, and created a slightly different one.

I wouldn't call the original intention fulfilled. Alice hadn't actually asked about Wasabi's personality. The database says `dropped`, not `done`.

But something useful happened: an old, vague intention was replaced by a more specific one, connected to an actual promise.

| id | content | source | status |
| --- | --- | --- | --- |
| 1 | Ask Michael how his son is feeling now that some time has passed | self_review | done |
| 2 | Check in about Wasabi's personality/antics | self_review | dropped |
| 3 | Follow up on whether Michael decided to upgrade his computer and what pushed him toward or away from an SSD | self_review | done |
| 4 | See how Michael feels now about the disagreement over honesty—whether the tension has settled or he's still annoyed | self_review | open |
| 5 | Ask the user to tell the story about Wasabi next time we talk | commitment | dropped |
| 6 | Ask Michael for the Wasabi story he promised | self_review | open |

Of course, there's a possible problem here. Two different mechanisms seem to be managing the same intentions, and we don't yet know whether they're coordinating properly or simply overwriting each other's work.

Still, the promise has survived the end of the conversation.

Now comes the interesting part.

Will Alice remember it next time, without me mentioning Wasabi at all?

## The Cat Strikes Back

Time for another conversation.

Alice has promised to remind me about Wasabi. Let's see whether she remembers.

But I won't make it easy. Instead of starting with something ordinary, I'll ask her a philosophical question about personal identity.

People change over decades, sometimes becoming almost unrecognizable. Yet we still consider them the same person. And what happens when two people spend years apart? Could they change so much that they no longer recognize each other?

A rather different topic from cat stories.

```plaintext
(alice7) mikhailburilov@MacBook-Pro-Mikhail alice7 % uv run alice-v2 chat
alice-v2 [switches: lessons,selfmodel,intentions] — db: data/v2/memory.db. /help for commands, /quit to exit.

you> Alice, I've been thinking about an interesting paradox. People grow older and change. Who they were at the beginning and who they become decades later might be almost unrecognizable, yet we still consider them the same person.Don't you think our sense of identity changes along with us? And if two people spend many years apart, could they eventually become so different that they no longer recognize each other as the people they once knew?

alice> Hey Michael! Before we go down the philosophy rabbit hole — you still owe me that Wasabi story, you know. Don't think I forgot.

But okay, this is actually a fun one. Yeah, I think you're onto something real. There's this classic thought experiment — the Ship of Theseus — where you replace every plank of a ship one by one, and at the end you ask "is it still the same ship?" People are kind of like that. Cells, beliefs, memories, habits — almost everything gets swapped out over decades. The guy at 20 and the guy at 60 might disagree about politics, music, what matters in life.

And yet there's usually a thread connecting them — memory, continuous experience, maybe just the story we tell ourselves about who we are. That's probably why we feel like the "same person" even if we're unrecognizable on paper.

As for two people growing apart over years — absolutely, that happens. You can love someone, lose touch, meet again, and realize you're now strangers wearing familiar faces. Not because either of you did anything wrong, just... different paths, different versions of yourselves.

What got you thinking about this — something specific, or just one of those afternoon brain-wander moments?

you> /end
session finished

you> /exit
```

And the very first thing Alice said was:

> Hey Michael! Before we go down the philosophy rabbit hole — you still owe me that Wasabi story, you know. Don't think I forgot.

Ha!

She remembered. And she brought it up herself, before answering my actual question.

Then she went on to discuss the Ship of Theseus and personal identity, apparently quite happy to switch from a cat story to philosophy.

But what happened to the intention?

After the session, Alice closed the previous one and created a new one:

> Finally get Michael's promised Wasabi story—he dodged it again into philosophy.

Now that's funny.

Not only did she remember the promise, she also decided that I was deliberately avoiding it.

Which, strictly speaking, isn't true. I was testing her memory and intentions. But she doesn't know that.

So we have another example of Alice interpreting my experimental behavior as something personal.

Still, I'm happy with the result.

An intention survived a session boundary, appeared spontaneously in an unrelated conversation, and remained relevant afterward. Alice even updated it based on what happened.

Does that mean she has genuine desires or goals? Of course not. We're observing a language model operating with persistent intentions, not proving anything about subjective experience.

But the mechanism works in this experiment.

And frankly, I didn't expect to be chased by my own AI over a cat story.

## What's Next?

We've now explored memory, lessons, beliefs, self-portraits, and intentions.

There are plenty of unanswered questions. How stable are these mechanisms over hundreds of conversations? Can intentions conflict? Can Alice abandon a goal without being explicitly told to? What happens when her beliefs change?

I don't know yet.

But I don't want to spend the next month testing every possible edge case of v2. There are more interesting things I want to build.

For now, Alice remembers what she wants to ask, and apparently doesn't give up easily.

That's enough for Day 4.

*To be continued.*