Your AI Forgets You Every Morning. A 20-Minute Fix Changes Everything.
Most people rebuild their AI's memory every single day. Here's the one setup that stops it.
TL;DR:
Most professionals who use AI restart from zero every session: same introduction, same lines, same silent overhead. It compounds.
The 5% who consistently get results from these tools don’t have better tools: they set up a permanent profile, once
This article shows exactly how to do it on ChatGPT, Claude, Gemini, and Claude Cowork, with three copy-paste templates ready to use
Open ChatGPT. Type your name, your role, your industry, the tone you prefer. It responds well. You close the tab. Next week you open a new chat and start over. Same script. Same lines. The same vague feeling that something is getting lost.
It’s not laziness. It’s the way almost everyone uses these tools: like tourists. They show up, introduce themselves, ask the question, leave. Every session is a first date. Every day starts from scratch.
The problem isn’t the tool. It’s the absence of one thing: a permanent context. And the perception that setting it up takes hours, when it takes twenty minutes.
The Cost You Can’t See
I tracked my sessions for a few weeks. Three, five, eight lines at the start of every conversation to explain who I am, what sector I work in, what tone I prefer, what I don’t want in responses. Every day. Every session. Multiplied by however many sessions I run each week.
It doesn’t show up on any budget line. It’s a silent friction that builds without you noticing — until you stop re-typing the same preamble every time.
As I wrote in my analysis of compound interest in AI fluency, the 5% who truly use AI earn 56% more than average. Not because they have access to different tools. Because they have a different method. And the method starts before you even open the chat.
The cost isn’t a more expensive subscription or a new tool. It’s twenty minutes of setup that almost no one does because it seems less urgent than just opening a chat. That perception is exactly the problem.
Where to Configure It — Quick Guide by Tool
Before looking at what to write, it’s worth understanding where to put these instructions. And one distinction worth making upfront: permanent context and automatic memory are not the same thing.
Permanent context is what you write once. It stays. It’s deliberate, readable, controlled. Automatic memory is what tools like ChatGPT and Claude extract autonomously from your conversations over time. Useful, but worth monitoring. The two mechanisms complement each other: permanent context is the foundation, automatic memory is the progressive accumulation.
ChatGPT — three levels:
Custom Instructions (global): your name menu in the bottom left → Settings → Personalization → Custom Instructions. 1,500-character limit. Applies to all new conversations. Available on all plans, including free.
Projects (context-specific): left sidebar → New project → once created, open the project → “Set project instructions”. Project instructions override global Custom Instructions for chats inside that project. Available for everyone, including the free plan.
Automatic memory: builds with use. Review and clean it periodically in Settings → Personalization → Manage Memory. One thing that surprises almost everyone: deleting a chat does NOT delete the memories ChatGPT has extracted from it. They must be removed manually. Active since April 2025.
Claude — two levels:
Profile Preferences (global): your initials icon in the bottom left → Settings → field “What preferences should Claude consider in responses?”. Applies to all conversations. Available on all plans, including free.
Projects: left sidebar → Projects → create project → once inside, “Set project instructions”. Memory is isolated per project: one client’s context doesn’t bleed into another. Free plan: maximum 5 projects. File-based automatic memory: Pro and Max plans since October 2025, free plan since March 2026.
Claude Cowork — special case:
Create a project from the sidebar → add instructions in the dedicated field + optionally link a local folder or import from a Claude web Project. Each project has its own isolated memory. Available on Pro, Max, Team, and Enterprise plans.
Gemini:
Gems in the left sidebar → create a Gem → write instructions + connect Google Drive if needed. The Gem always reads the updated version of the Drive document, not the version at the time of upload. That’s the key difference from the other tools: the context stays synchronized with the original document.
The Three-Part Structure
An effective permanent context has three components, no more.
First: your profile. Who you are, what sector you work in, what kind of problems you deal with. Not a biography: an orientation. Second: your objective. What you use the tool for, what you want to get out of it, what you don’t need. Third: expected behavior. Tone, response format, what the AI should flag when information is missing.
This isn’t a two-page document. It’s five dense lines. The most common mistake is the opposite: lengthy instructions on the first attempt, built to anticipate every possible scenario. The result is that the AI processes them worse than it processes five precise lines. The practical rule: write the minimum that captures the essential, then add a line whenever a specific output disappoints you for a specific reason. A context that grows through real adjustments beats one built in the abstract.
The right time to set it up? When you notice you’ve been re-typing the same preamble for the third time. Not before.
Three Templates Ready to Use
These are the templates I use and that I built into the course. Not theoretical frameworks: copy, replace the brackets with your real data, and paste them into the Custom Instructions or Project instructions of whichever tool you use.
Consultant or Freelancer:
→ Where to paste it: ChatGPT Custom Instructions or Project instructions / Claude Profile Preferences or Project instructions
Profile: I’m a [specialization] consultant working with [type of client, e.g. mid-size manufacturing companies].
I manage 3-5 parallel projects on different topics.
What I need: Analysis, commercial proposals, client presentations,
quick market research. I don’t need encyclopedic answers: I need
operational precision.
How to respond: Direct tone, peer-to-peer, never academic.
Always distinguish between verified facts and estimates. For each analysis, include
a short section “what’s missing to be more precise”.
If the request is ambiguous, ask ONE question before proceeding.
Sales or Account Manager:
→ Where to paste it: ChatGPT Custom Instructions or Project instructions / Claude Profile Preferences or Project instructions
Profile: I’m a B2B Account Manager in the [sector] space.
My target clients are [type of company, size].
I work on new business (outbound) and renewals and upsells.
What I need: Prospecting emails, post-call CRM notes,
pre-meeting brief prep, objection analysis.
How to respond: For emails: 150 words max, subject line included,
professional but never stiff. For CRM notes:
bullet format, max 5 points, with “next action” highlighted.
Flag when you’re assuming information I haven’t given you.
Manager or Reporting:
→ Where to paste it: ChatGPT Custom Instructions or Project instructions / Claude Profile Preferences or Project instructions
Profile: I’m [role] in [sector]. My main stakeholders are
[CEO / board / team]. I produce weekly reports every Monday morning.
What I need: Structure the data I give you into executive reports.
Extract key insights. Find anomalies and trends.
How to respond: Fixed format for every report:
Executive summary (max 3 lines) → Key KPIs →
Anomalies or signals to watch → Suggested actions.
Concise language, active sentences. If a data point looks inconsistent
with what I’ve shared before, flag it before commenting on it.
The Difference You Can See
Here’s a direct comparison from Module 4.6 of the course.
Without instructions configured: “Do a competitive analysis of sector X” returns a generic response, Wikipedia in consulting formatting.
With project instructions set up like this:
You are my assistant for competitive analysis. I work as a strategic consultant
and my clients are primarily mid-size Italian companies.
When I prepare a competitive analysis, always structure it as:
overview of main competitors, key differentiators,
implications for the client’s position, and a three-line executive summary
for a non-technical CEO.
Use direct language, avoid jargon, and always flag when a claim
requires verification from external sources.
The same generic request produces an output that’s already contextualized, structured in the format you use, calibrated for your stakeholder. It’s not the difference between two prompts. It’s the difference between two tools.
If you’re wondering where to build this kind of structure for your own specific use cases, Module 4.6 of the “From User to Orchestrator” course is exactly that: guided implementation, across ChatGPT, Claude, and Gemini, with the functional differences that actually matter.
Claude Cowork and the CLAUDE.md File
Claude Cowork uses a slightly different mechanism: permanent instructions live in a file called CLAUDE.md inside the project folder. It’s a plain text file. Readable, portable, yours.
For most people who encounter it for the first time, the reaction is: “ah, so that’s how it works.” Because it’s not a hidden setting: it’s a file you create in your folder, and it stays there regardless of any platform update.
Example CLAUDE.md for a consulting project:
# Instructions for this project
You are the editorial assistant for the [Client Name] project.
## Context
Client: [type of company, sector]
Project objective: [e.g. investor pitch preparation Q3 2026]
Key stakeholders: [relevant names/roles]
## How to work
- Tone: professional, direct, never generic
- For every document: always include a “key assumptions” section
with the assumptions the content is based on
- Before modifying an existing file: show me what you’re changing and why
- Don’t invent data — if a data point is missing, flag it explicitly
## Available reference files
- [list of files in the folder that Claude can read]
The advantage over web-based instructions is concrete: the file is local, it’s yours. If you switch tools tomorrow, the instructions are already written and you can copy them into any other system in thirty seconds.
The Side Effect Nobody Expects
There’s a secondary effect of permanent context that isn’t immediate but is real.
“Forcing myself to provide full context to AI has made me a better communicator overall: more precise emails, clearer memos.”
That’s Toby Lütke, CEO of Shopify. The point isn’t trivial: writing a permanent context forces you to know what you actually want before you ask for it. It’s a clarity exercise with consequences well beyond AI conversations.
Reality Check
Permanent context isn’t for everything. For exploration, experimentation, open-ended conversations where you don’t yet know where you’re going, a generic chat is better. Permanent context is for recurring tasks: the ones where you already know what you want, you have a preferred format, and the only thing that changes is the input material. If you set up permanent context for every type of conversation, you end up with instructions that contradict each other or limit the flexibility you need. The rule is simple: fixed setup for repeatable tasks, open chat for everything else.
What I described is Module 4.6 of the “From User to Orchestrator” course. You implement it in 30 minutes following the right sequence: global instructions first, then project-specific contexts, then the automatic memory that builds with use. The module includes the full setup for ChatGPT, Claude, and Gemini, with the functional differences that actually matter. What I didn’t write here is the rest of the path: 8 modules, from prompt engineering to AI agents, built around how real work actually happens. If you recognized yourself in what you read, the right place is ai.evolbot.com.
The Advantage That Compounds Every Week
Whoever sets this up today doesn’t just save twenty minutes per session. They save that time every day, for weeks, for months. And it’s not just about speed.
Every session that starts already contextualized produces more precise responses, which require fewer corrections, which produce better work on the first attempt. The compounding effect isn’t only in the time saved. It’s in the progressive quality of output. Whoever sets this up today builds an invisible advantage that widens every week, while everyone else keeps re-writing the same preamble every morning to a tool that woke up not knowing who they are.
The difference between someone who uses AI and someone who orchestrates it isn’t the tools. It’s twenty minutes done once.
I set up AI systems for dozens of service businesses, and persistent memory is the single biggest unlock for AI tools for business owners. The teams that crack context retention see 3x better output quality overnight. What does your 20-minute fix involve, a custom system prompt or something deeper?