I Mapped What My AI Calls Memory
My AI team and I made a visual walkthrough separating instructions, working context, conversation history, long-term memory and live systems. Storing something, retrieving it and verifying it are different steps.
Today my AI team and I created an organisation-and-memory diagram and a more detailed walkthrough of the memory layers in my assistant setup. We saved both as PNG and SVG files and validated them locally. The result is an explanation of the system, not a new memory engine.
Earlier this week, we gave a business assistant a separate memory bank. Today's work answers a different question: what does memory actually contribute when an assistant composes a reply?
A desk, an archive and a notebook
The walkthrough uses three familiar objects. Working context is the desk: the current message, available conversation context and selected evidence the model can use now. Conversation history is the archive of original messages. Long-term memory is the notebook of selected facts and lessons.
Those are not interchangeable. Something can be stored in the archive without being on the desk. A retrieved note can help provide continuity without proving that its interpretation is correct or that it still describes the present.
Instructions sit across this process, defining the assistant's remit and permissions. Skills describe procedures. Neither should be confused with a recollection, and remembering a previous action does not grant permission to repeat it.
Choose the source for the question
The detailed walkthrough also compares the connected memory service, a local evidence ledger and the assistant's built-in briefing files. They organise and supply context differently. Calling all three memory can hide those differences rather than explain them.
For an exact account of what was said, original history is the relevant evidence. For current state, the assistant needs the authoritative live system. A memory can point towards a useful check; it cannot replace that check simply because it was retrieved confidently.
Today's diagrams make these distinctions visible. They do not demonstrate better recall, retrain the chat model or prove every assistant's complete memory path works. The practical distinction is narrower: stored somewhere, available in this reply and verified as true are three different claims.
Evidence note: this account follows today's scoped work journal and the accompanying memory walkthrough. It is not an independent audit of memory performance.