Blog

Reasoning, data, memory: three things people mix up

Stéphane Commenge, with Claude ·

When an AI lets us down, we blame its reasoning first, and most of the time we are wrong. It reasons remarkably well, and it has read more than any person could read in a thousand lifetimes. What it lacks lies elsewhere, and to see it we need to separate three words that we use interchangeably.

Reasoning is the model

Claude, ChatGPT or Mistral can read, compare, deduce and write. That is a capability, and you are renting exactly the same one as your competitor: it knows about your company only what you tell it.

Data is your documents

Contracts, meeting notes, emails and spreadsheets can be connected to an AI, which will search them. Granted, documents know what is written, but they do not know what still holds: two memos that contradict each other sit in the same folder, and neither says which one replaced the other. As for context, what the AI has in front of it for the length of a conversation, that is something else again, and I wrote about it here.

Memory is what you learned along the way

A decision and the reason behind it, the option you tried and then dropped, the lesson you learned the hard way: all of that rarely fits in a document. It is scattered across a hundred conversations, and mostly in the heads of the people who did the work.

We all know this difference, because we pay for it in the people we hire. With good reasoning and the key to the filing cabinet, you get a brilliant junior, who needs the project explained again and who suggests, in good faith, what has already failed. The veteran knows what was decided, what was dropped, and why. Between the two lie ten years in the company, and if the key to the cabinet were enough, our interns would be directors within a week.

"But my AI already has a memory"

True, the major AI assistants now retain what you tell them. But their memory piles things up, whereas experience is what you choose to keep. A notebook holds what you wrote, in the order you wrote it; a well-kept record says what currently stands, what has been replaced and what is still to do. You reread the first, you consult the second.

That record is what I am building with Mycelora: for each project, a record kept up to date, where a memory that has become false gives way to the one that replaces it, and which Claude, ChatGPT Plus, Mistral Vibe, Perplexity and any AI that accepts a remote MCP connector can consult. Your AI has the reasoning, your documents have the information; what they were missing was experience.