You can hand an AI a detailed prompt and still get back something generic. It hits the word count, the format and the tone, and still misses why you kept one paragraph, cut a hook you had already used twice, or refused to make a claim stronger than the evidence behind it.

The prompt did its job. It had your instruction. It did not have your working record (context).

That gap is now the whole game, because the tools have stopped competing on reasoning alone. OpenAI shipped a memory system called Dreaming in June that reads across your past chats in the background and keeps its picture of you current on its own. Anthropic put memory on every Claude plan in March. Google’s Personal Intelligence connects Gemini to Gmail, Drive and Maps. Microsoft finished rolling out Copilot Memory in May. Four of the largest companies in software made the same bet inside five months.

One Topic: Your AI Assistant Needs a Better Record of You

A prompt captures the task, a record captures the judgement

A prompt explains what you want right now. A record shows how you have decided similar things before.

For a newsletter, “keep the main section under 1,000 words” is a useful instruction. Your corrections teach far more. They show that you cut manufactured urgency, distrust unsupported authority, and kill any social post that gives away the article. Save the reason alongside each correction and the AI learns the standard behind the fix rather than the fix alone.

Anthropic calls this wider practice context engineering, and defines it as curating the whole set of tokens a model sees during inference, including everything that lands there from outside the prompt. OpenAI is building toward the same thing from the product side. When it released a memory preview for Codex, it described what gets remembered as personal preferences, corrections, and information that took time to gather. That last category is the one people skip, and it is the one that stops every task from restarting at zero.

Why more is not the same as better

The obvious move is to feed it everything. That fails in a measurable way.

Chroma Research tested 18 frontier models, including GPT-4.1, Claude 4, Gemini 2.5 and Qwen3, holding task difficulty constant and varying only the input length. Every model got less reliable as the input grew, even on simple retrieval. They called it context rot, and the important part is that it starts well before the window is full. A 200,000 token window can degrade noticeably at 50,000.

So volume is the wrong target. The question is what earns a place.

The five things worth keeping

1. Current priorities, with a date and a review point. Undated priorities quietly become permanent rules long after they stop being true.
2. Decisions, including what you rejected and why. Most people log the choice. Almost nobody logs the reasoning, which is the only part that transfers to a situation you never described.
3. Corrections that show up more than once. A repeated correction is a rule you have not written down yet. The Manus team found the same principle building their agent and published it as “keep the wrong stuff in”: hiding failed actions and error traces stops the model learning from them, while leaving the failure visible pushes it away from repeating the mistake.
4. Approved examples that show what good looks like. One accepted piece with a note on why it worked beats a page of adjectives about your tone.
5. Work traces, from first attempt to accepted result. The final document shows the destination. The drafts show how you got there. An AI that can see you removed an inflated statistic, cut a slow opening and rewrote a salesy ending has watched your taste operate. Stack Overflow’s 2024 survey found 63% of developers named missing context about their codebase, architecture and institutional knowledge as a top concern with AI at work. That gap is not specific to code.

Voice notes help here, because talking is faster than writing a decision log. Two minutes after a review on what changed and why is enough. The recording should not become permanent memory, though. Turn it into one dated line you can check later.

Where to put it

Start with one workflow you already repeat every week. A full second brain built on day one becomes another system to maintain.

Four plain Markdown files do the job: Profile for stable preferences, Decisions for dated choices and reasons, Corrections for repeated mistakes and the rule behind each, Examples for approved work with a note on why it worked.

At the end of a task, ask the assistant to propose no more than three lessons, then review before saving. Once a week, merge duplicates and delete what is no longer true. OpenAI’s own cookbook on context personalisation recommends exactly this shape: take notes during a session, consolidate them into a smaller memory afterwards, and prune anything stale, duplicated or low signal.

Keep it in Markdown rather than locked inside one tool. If you switch assistants next year, plain text moves with you.

The boundary

Passwords, private messages, confidential company material, financial and health details, and anything sensitive about other people stay out. Personal and work records stay separate. The UK’s data protection regulator advises organisations to limit personal data to what the purpose actually requires and delete what is no longer needed, which is a sound rule for one person’s setup too.

Decide why you are saving something before you save it.

The models are close enough in ability now that picking between them rarely decides your output. What decides it is how much of your actual working life the thing has seen.

Most people are still trying to write a better prompt. The quieter advantage is a better record.

Start with the corrections from your last ten tasks. Add the reason behind each one, strip anything sensitive, and run that record against a single repeated workflow for a month. You will find out fast which details improve the work and which ones were only noise.

Interested in travel or photography, read last week’s LensLetter newsletter about the tripod I’ve been avoiding for 15 years.

Read last week’s JustDraft about ChatGPT Codex Computer and Browser Use cases.

Two Quotes to Inspire

Judgment is not what you decided. It is the list of things you turned down and the reasons you had for each one.

Anyone can copy your process. What takes years to copy is the record of everything you tried that did not work.

One prompt to steal

Use this the moment you finish rewriting something the AI got wrong. It converts a one-off fix into a standing rule.

"I just rewrote your draft. Compare your version against mine and tell me what changed. Then write the underlying rule as one sentence I could hand to someone new. State the standing preference behind the correction rather than the specific error you made this time. If you cannot tell whether it was a rule or a one-off, ask me."

Save the rules somewhere durable. In about six weeks you will have a working style guide you never sat down to write.