Michael Kraewing · Interim Executive Reference · one page

Tools · The Memory Sheet

Four properties of the machine you are talking to.

An AI forgets without any sense that something is missing. That single fact explains most of what goes wrong — and each of the four properties below has a practical consequence you can apply tomorrow.

Free to use · applies to any current AI assistant

The starting point

It forgets — and doesn't notice.

When you forget, an edge remains: the name on the tip of your tongue, the sense that something was there. An AI keeps nothing — no edge, no gap, and no feeling of a gap. It begins every session fully trained and entirely without history.

Which is why it is never at a loss. It doesn't hesitate, it doesn't say "hold on, there was something". It answers fluently, immediately, in complete sentences. A person with that amnesia would be conspicuous. A machine simply sounds competent — and that is the property most people underestimate.

The four properties

What is actually going on underneath.

None of these is a defect to be fixed. They are structural, they apply to every current model, and each one has a direct consequence for how you brief and check.

Reading and writing

Fragments, not letters

Your text arrives all at once, as a surface; the answer is assembled piece by piece. The model thinks in fragments of words, not characters — it can write a page about a word and fail to count its letters.

Numbers

Shapes, not values

A string of digits carries no place value. The model recognises the silhouette of a number, not its magnitude. An invented figure looks exactly like a retrieved one.

Attention

Lost in the middle

Long input is not read evenly. Beginnings and endings stay vivid; the middle fades. Put something important in the middle and you have not emphasised it — you have hidden it.

Output

A distribution, not a truth

Each pass ends in a probability distribution. Ask the same question twice and you get two answers, both equally confident. Reliability comes from the setup, not from the model trying harder.

In one picture

Where your instruction lands.

Attention across a long instruction: high at the beginning, low in the middle, high again at the end. ATTENTION ACROSS A LONG INSTRUCTION Beginning Middle End read closely quietly skimmed read closely
The same sentence is worth more at the top or the bottom of a brief than in the middle of it. Constraints, exclusions and acceptance criteria belong where attention is highest — not buried in the body of the text.

What follows

Four habits, one per property.

  1. Have it fetch, never generate

    Every figure, date and name that matters comes from a named source — a file, a document, a search result. "From memory" is not a source. If it cannot be fetched, the honest answer is "unknown".

  2. Put the critical part first or last

    Constraints and acceptance criteria at the top or the very bottom of the brief. If something absolutely must not be missed, it belongs in both places — repetition costs nothing.

  3. Distrust the garnish, not the claim

    The main point has usually been researched. What tends to be invented is the incidental precision around it: a date in a subordinate clause, a size in passing, a casual "since". Check those.

  4. Write the state into files, not into the chat

    A conversation is volatile; a small set of maintained files is not. Read them at the start of every session, write results back at the end. That is what turns a brilliant stranger into the same colleague as yesterday.

In short

The difference between a tool and a colleague isn't intelligence. It's a past — and an AI's past doesn't sit inside the model. It sits in the files you keep for it.

The most valuable thing in those files is not the record of what went well. It is the dated, unvarnished record of what went wrong.