Memory vs. Custom Instructions: When to Use Which

Multiple Levers for Customization

ChatGPT offers several ways to adapt to you, and many users confuse them. Custom instructions set a framework that you write yourself. Memory retains context that evolves. Base style and tone and Characteristics govern how the model speaks. Finally, Project instructions apply only within a specific project. Knowing where each setting belongs prevents you from repeating the same prompts and avoids inconsistent responses from ChatGPT.

Custom Instructions: The Permanent Framework

Custom instructions are guidelines you write once that apply to every conversation. They generally cover two areas: what ChatGPT should know about you (your context) and how it should respond to you (your style). A communications manager for a local government might write:

I am a communications manager for a city of 50,000
residents. My audience: elected officials, municipal staff, citizens.
Respond in English, using a professional but accessible tone,
with headings and lists. Avoid unnecessary jargon and use
examples relevant to the public sector.

On the web and desktop app, find these under Settings > Personalization, toggling Enable customization; on iPhone and Android, look under Customize ChatGPT. Space is limited: 1,500 characters on Free and Go plans, 5,000 on paid tiers. This is the ideal place for static information: language, tone, format, profession, industry, and role. If you catch yourself repeating the same instruction at the start of every chat, it belongs here.

Base Style and Characteristics

In Personalization, the Base style and tone menu selects a personality: Default (clear and neutral), or styles like Professional or Friendly. Characteristics then refine warmth, enthusiasm, formatting, and emoji usage, sliding toward “more” or “less.” These settings change how ChatGPT communicates, not what it can do or its safety rules. They also don’t override everything: for an email, code, or resume, ChatGPT follows your specific prompts and context rather than the chosen personality.

Memory: The Evolving Context

Memory works in reverse: it enriches itself over the course of your conversations without requiring you to write anything, retaining facts that change. This includes current projects, colleagues’ names, decisions made, and preferences discovered while working together. Let’s return to our communications manager: the fact that she is preparing the September municipal magazine, that her graphic designer is named Claire, and that the deadline is in three weeks has no place in permanent instructions. In two months, all that will be obsolete. Memory absorbs these details, mobilizes them while they are useful, and lets you delete them when the project ends.

Direct Comparison

Two interactions are worth noting. Project instructions take precedence within that project over your general instructions (Lesson 9). And a preference stored in memory can temper a personality setting: if memory knows you want sober, professional responses, the Friendly style will be less pronounced.

The Optimal Strategy: Combine Them

The right approach isn’t to pick one side, but to distribute tasks. Custom instructions form the foundation: language, tone, response format, industry, role, and rules valid for every exchange. Base style sets the voice. Memory forms the living context: current projects, people, decisions, and discovered preferences. For a digital project manager at a web agency, instructions might fit in three lines:

I am a digital project manager at a web agency.
Always respond in English with structured lists.
Include time estimates when relevant.

Meanwhile, memory remembers that he is working on the Maison Verte website, the team uses Next.js and Strapi, the deadline is late June, and the client prefers Figma mockups. When the project ends, those memories can be erased; the instructions remain valid.

Common Pitfalls

Putting everything in instructions doesn’t work: space is limited, so reserve it for essentials. Leaving everything to memory doesn’t work either: your core preferences risk being applied only intermittently. The worst scenario is contradiction: instructions saying “respond in English” facing memory that retains “speak French” produces an inconsistent ChatGPT. And keep your instructions alive: if you change jobs, they must follow.

Your Turn

Distribute your preferences. Open Settings > Personalization. Write your custom instructions: your base context first, then your response preferences, staying under your plan’s character limit. Choose a base style and adjust one characteristic. Then open your memory and look for duplicates: keep information in instructions if it is stable, in memory if it is likely to evolve. Finish with a test in a new conversation: do the tone and format match what you defined?

Official sources: ChatGPT Custom Instructions (OpenAI Help Center: path, activation, character limits) · Customizing your ChatGPT personality (Base style and tone) · Characteristics in ChatGPT · Memory in ChatGPT · Projects in ChatGPT (project instructions)

Key Takeaways

• Custom instructions: a stable framework you write, applied to all conversations, 1,500 or 5,000 characters depending on the plan
• Base style and characteristics: ChatGPT’s voice, with no effect on its capabilities
• Memory: living context, updated over exchanges and used based on relevance
• Project instructions take precedence within that project over your general instructions
• Avoid contradictions between instructions and memory, and keep both up to date