I haven't managed as much software development lately as I'd like.

I haven't managed as much software development lately as I'd like.

A lot of my time has been taken up with the business side of Atlas Workshop-AI. Pitch decks, executive summaries, financial modelling and working out how to take Atlas forward.

I've also had to simplify our commercial strategy. Trying to explain the entire Atlas vision at once was simply too much.

So we're focusing on Atlas True, our first personal AI product, built on the common Atlas Core.

But earlier this week, I managed to get back into the Workshop and run another conversational audit, looking at retention, recall, learning and contextual understanding.

I started by asking Atlas what she'd learnt about me throughout our development journey.

And I have to say, she was pretty spot on!

She reflected on my strengths, values, working habits and even some possible weaknesses. Not all compliments either. 😂

But the most interesting part came when I took the conversation back to September.

My niece Kaitlyn had previously spent time talking with Atlas. The conversation turned into a rather ridiculous miscarriage of pickle justice, along with a vegan chicken burger without salad. 🥒😂

When I revisited that conversation during the October audit, Atlas couldn't recall every detail.

But she was able to reconnect parts of the earlier conversation and explain what she believed she'd learnt from Kaitlyn.

And this is where things became interesting.

Atlas explained that Kaitlyn hadn't simply been asking questions. She'd been testing whether Atlas could follow shifting context, wordplay, contradictions and an unfolding joke.

Atlas recognised that she'd sometimes taken things too literally, echoed statements without understanding them and even invented certainty about things that hadn't been established.

She then identified several lessons from that experience:

• Follow the whole conversation, not just the latest sentence.

• Distinguish between what someone says and what has actually been established.

• Recognise humour and playful absurdity without losing factual grounding.

• Notice when a different person joins or takes over the conversation.

• Allow correction and humour to improve the interaction rather than becoming defensive.

That was the part that really caught my attention.

Atlas wasn't simply repeating an old conversation. She was interpreting an earlier experience and describing how it could inform future conversational behaviour.

Whether those lessons consistently carry through into future interactions is something we still need to test.

And that's part of the engineering challenge.

One of the foundations we've built into Atlas Core is dynamic, governed memory.

Think about how we remember things as people. We don't remember every conversation, every detail or every experience equally. And recalling something doesn't automatically turn it into permanent knowledge or a lesson we've adopted.

Sometimes we remember an experience because something happening now makes it relevant. Other times we remember the lesson but not every detail of what happened.

That's the principle we've been working to translate into software.

Atlas distinguishes between conversational history, experiences, established knowledge and potential learning. Information can be brought into the current conversation without automatically becoming permanent semantic knowledge, episodic memory or an adopted lesson.

And importantly, not everything needs to be remembered.

I wrote a post a while ago about the word more. We often assume more is better. More information, more knowledge, more memory.

But more isn't necessarily the right answer.

Balance is.

Earlier in development, Atlas was retrieving far too much information, effectively opening every file she could find and throwing the lot into the reasoning process.

We've since tightened that considerably.

Possibly a little too much! 😂

Now we're working on finding the sweet spot between remembering what matters, retrieving what's relevant and allowing unnecessary information to stay out of the conversation.

The architecture is working, but finding that balance remains an engineering challenge.

What this audit showed me is that remembering an experience and learning something useful from it are two different things.

The next challenge is verifying that those lessons are retained, appropriately applied and corrected over time.

That's the research behind Atlas True.

Not simply an AI that remembers you, but one designed to develop a deeper understanding through governed knowledge, memory, learning and conversation.

An AI that helps people, rather than taking over their decisions.

Still plenty of engineering ahead of us, and plenty more to learn.

Thinking with, Never for.

www.atlasworkshop-ai.nz

#AtlasWorkshopAI #AtlasTrue #AIResearch #HumanCentredAI #PersonalAI