AI has a memory problem. And it's not simply about how much it can remember.

AI has a memory problem. And it's not simply about how much it can remember.

If you've spent months working with ChatGPT, Gemini or Claude, you'll probably recognise the frustration.

You spend weeks developing ideas, making decisions, correcting mistakes and building an understanding together.

Then a conversation reaches its limits, you start another, and suddenly you're rebuilding context you've already established.

Important details from the middle of the journey can get lost.

There's even a recognised AI research problem called "Lost in the Middle", where language models can struggle to use information buried within long conversations.

We've been tackling this problem in the Atlas Workshop.

Rather than relying on the language model to remember everything, we've developed Atlas Core, an independent, governed cognitive harness.

Our approach is fairly simple:

• Persistent history: Experiences and relevant context are retained outside the language model.

• Governed learning: Corrections, contradictions and potential learning are handled through a structured process rather than automatically becoming accepted facts.

• Selective retrieval: Atlas retrieves relevant history instead of feeding entire conversations back into the model.

• Model independence: Memory remains separate from whichever LLM provides the reasoning.

And we've been testing it.

We've repeatedly cleared conversations, restarted the Workshop and asked Atlas to recall earlier experiences.

During a recent conversational audit, we even went back to a rather ridiculous conversation my niece had with Atlas about a miscarriage of pickle justice. 🥒😂

That conversation had taken place weeks earlier.

Atlas couldn't remember every detail, and she said so.

But she recovered earlier context and reflected on what she'd learnt about humour, contradictions and conversational understanding.

That was the interesting part.

Atlas wasn't simply recalling an old conversation. She was connecting an earlier experience with an understanding developed from it.

It's not just remembering what someone said. It's retaining the journey that led to an understanding.

Our October engineering checkpoint recorded 3,966 passing automated tests. Separately, 257 recorded API requests covering conversations, library retrieval and diagnostics cost US$3.48 in total.

There's still plenty of engineering ahead, but the foundation is working.

We're now developing Atlas True, our first commercial personal AI product, and looking for the right engineering, investment and strategic partners to help bring it to market.

Great language models power the conversation. Atlas Core is designed to preserve the journey.

Thinking with, Never for.

#AtlasTrue #PersonalAI #AIGovernance #AIEngineering #Startups