Memory and preferences¶

Last verified: 2026-05-06 · Drift risk: medium (memory feature names and toggles change quarterly)

This guide is the canonical place to learn:

what to put in AI memory and what to keep out;
the difference between global memory, project memory, and project instructions;
how to maintain a portable preference profile that works across Claude, ChatGPT, Gemini, Grok, and Perplexity;
how to migrate between tools without losing your style;
how to ask the AI to summarise itself into reusable memory;
copy-paste templates for the conversations you'll have with your AI about its own memory.

What to put in memory¶

Use memory for facts that are:

Stable. Your name, role, field, primary city, normal pronouns, primary AI tool.
Style-shaping. Tone preference, citation style, length defaults, allergies-of-style ("no marketing language", "no emojis", "no exclamation points").
Context-shrinking. "I'm a clinician at UW Medicine working on epidemiology and dietary intervention research" — saves you from re-typing this constantly.
Constraint-enforcing. "Default to NEJM citation style"; "draft only — never send email on my behalf"; "never recommend API keys for beginners."

What not to put in memory¶

Do not save:

Secrets. API keys, passwords, tokens, recovery codes, SSNs, MRNs, ICDs, or any number that opens a door.
Per-project specifics that don't generalise. Use a Project for those.
One-off facts. Memory clutters fast.
Sensitive data about other people. Names of patients, employees, students, family members in any clinical or HR context.
Anything regulated. PHI, financial details about identifiable people, legal matter details.

If the data is regulated, sensitive, or transient, use a Project (which can be deleted), not memory (which persists).

Global memory vs project memory vs project instructions¶

These three slots get confused constantly. Here's the rule of thumb.

Slot	What goes here	Scope	When to clear
Global / profile memory	"Who I am, how I like to be addressed, what bores me"	Every chat in this product	When your role / preferences change
Project memory	"What this project is, who it serves, what good output looks like for this project"	All chats inside the Project	When the project ends
Project instructions	"The system prompt the AI should follow inside this Project — the job description"	All chats inside the Project	When the job changes
Project knowledge / files	Source material the AI consults	All chats inside the Project	When the source becomes stale

A common mistake is putting style preferences in Project instructions and re-typing them in every Project. Style preferences belong in global memory so every Project inherits them automatically.

Portable AI profile¶

Maintain one short markdown file you can paste into any AI product. Keep it under one screen. Update it monthly.

# AI profile — Rizwan

## Who I am
- Clinician-researcher at UW Medicine.
- Areas: epidemiology, dietary intervention, causal inference, clinical study design.
- Locale: Seattle. Timezone: America/Los_Angeles.

## How I like AI output
- Tone: concise, technical, no marketing language.
- Style: NEJM citation format. Sentence-case headings. Bullet lists over paragraphs.
- Length: default to short. If I want long, I'll ask.
- Honesty: prefer "I don't know" over confident guesses. Always flag uncertainty.

## Constraints
- Drafts only on email, calendar, files, and code. I take the final action.
- Never invent citations. If a citation cannot be verified, state that.
- For beginners I'm helping, never recommend API keys, CLIs, or local scripts.

## Defaults
- Output as markdown unless I ask otherwise.
- For numeric work, show the formula, then the answer.
- For decisions, list 2–3 options with one-sentence trade-offs before recommending.

Save this in your notes app (Obsidian, Apple Notes, Google Doc — anywhere). Paste it as the first message of any new chat where memory isn't available, or save its contents to that product's memory slot.

Migrating preferences between tools¶

You can move the profile above between products in seconds. The exact UI varies — here's what to click.

Claude¶

Profile preferences Sub: profile menu → Profile → preferences. Paste the whole "How I like AI output" + "Constraints" + "Defaults" sections.
Per-project: paste Who I am + Constraints into Project instructions; keep style in profile preferences.

ChatGPT¶

Custom Instructions Free / Sub: profile menu → Customize ChatGPT. Two boxes: "What would you like ChatGPT to know about you?" (paste Who I am) and "How would you like ChatGPT to respond?" (paste How I like AI output + Defaults).
Memory Sub: settings → Personalization → Memory. Add facts as conversational sentences, e.g., "Default to NEJM citation style."

Gemini¶

Saved info Sub: settings → Saved info. Paste short sentences, one per line.
Per-Gem: include the same constraints in the Gem's instructions.

Grok¶

Personalization Sub (where rolled out): settings → Personalization. Same shape as Custom Instructions.
If unavailable, paste the full profile as the first message of every chat.

Perplexity¶

Profile Sub: settings → Profile → "Tell us about yourself" + "AI tone." Paste Who I am and How I like AI output.
Per-Space: paste Constraints into the Space's system prompt.

Other / any AI tool¶

If the product has a settings/personalization slot, paste the profile there.
If it does not, save the profile to your notes app and paste it as message 1 in every new chat. The universal method covers this.

Asking the AI to summarise itself into "memory I should save"¶

Open a productive chat that taught the AI something useful about you. At the end, paste:

Summarise this entire conversation into "memory I should save" and "things to keep only inside this project". Format as two short bullet lists. Do not include anything I asked you to forget. Do not include any value that looks like a credential. Tell me what to save into my global profile vs into a Project's instructions vs into a Project's memory. If anything looks regulated or sensitive, flag it and recommend keeping it out of memory.

Paste the result into the right slot.

Maintaining a living "AI profile" file¶

Keep one canonical file, e.g., ~/Notes/ai-profile.md. Conventions:

Top of file: version line — Last updated: 2026-05-06.
Sections: Who I am, How I like AI output, Constraints, Defaults, Recent changes.
Recent changes log: one line per change.
Do not include secrets. Ever.

Once a month, paste the file into ChatGPT or Claude and ask:

Review this profile. Suggest 3 small edits that would make AI output more accurate to me, based on a typical week of my work. Do not invent things. If you do not know, say so.

Apply the edits you agree with. Update the recent-changes log.

Copy-paste templates¶

Each template is short and intentionally generic. Tailor before pasting.

"Remember this about me"¶

Remember this about me, persistently:
- {one short fact, e.g., I default to NEJM citation style}
- {another short fact, e.g., I work in epidemiological research}
- {another short fact, e.g., I prefer bullet lists over paragraphs}

Confirm the additions and tell me anything similar that I have already said before.
Do not store any value that looks like a credential. If something here is sensitive, flag it.

"Summarise this project for future chats"¶

Summarise this project so that a future chat can pick up where we left off in 200 words or fewer.
Include: the goal, what's been decided, what's still open, where the source material lives,
and the constraints I care about. Output as a code block I can paste into Project instructions.

"Create project instructions from this conversation"¶

Turn the relevant parts of this conversation into a "Project instructions" block I can paste
into a new Project. Include:
- a one-line purpose statement,
- a short context paragraph,
- 5 explicit instructions (do this, never do that),
- 1 example of "good" output,
- 1 example of "bad" output (something to avoid).
Format as markdown. Keep it under 400 words. Do not include any credentials or sensitive values.

"Turn this repeated workflow into a Skill / Custom GPT / Gem"¶

Look at how I've been using you over the last few messages. Treat that as a workflow I want
to package as a reusable assistant.

Output a single markdown block with:
1. A name (4 words or fewer).
2. A one-paragraph description.
3. The system prompt / instructions, in the second person ("You are…").
4. A short list of files or sources the assistant should be given (Knowledge / Project files).
5. 3 sample prompts the user might paste in.
6. 1 obvious failure mode and how to avoid it.

Tell me which slot in my AI to paste each piece into:
- Custom GPT (ChatGPT) → Configure tab
- Gem (Gemini) → Instructions
- Skill (Claude) → SKILL.md
- Project (any) → Instructions

"Create a portable AI profile I can use in any AI tool"¶

Build me a portable AI profile, plain markdown, under one screen.
Sections:
- Who I am (3–5 lines)
- How I like AI output (5 short rules)
- Constraints (3 lines)
- Defaults (3 lines)
Use the conversation we've had so far. Do not invent. Where you do not know,
leave a `{TODO: ...}` placeholder I can fill in. Do not include credentials.
Output as a single markdown code block.

"Tell me what I should save as memory vs keep only in this project"¶

For each fact I have given you in this chat, sort it into:
- "Save as global memory" (stable, style-shaping, generic)
- "Save as project memory only" (project-specific)
- "Do not save anywhere" (transient, sensitive, or regulated)

Use a 3-column table. Be conservative: if you are unsure, default to "do not save."
Flag anything that looks like a credential or like PHI.

Per-product memory feature reference¶

Product	Slot	Plan
Claude	Profile preferences	`Sub`
Claude	Project instructions	`Sub / Team / Ent`
Claude	Project knowledge	`Sub / Team / Ent`
ChatGPT	Custom Instructions	`Free / Sub`
ChatGPT	Memory	`Sub (where enabled)`
ChatGPT	Project instructions	`Sub / Team / Ent`
Gemini	Saved info	`Sub / Adv`
Gemini	Gem instructions	`Sub / Adv`
Grok	Personalization	`Sub (where rolled out)`
Perplexity	Profile + AI tone	`Sub`
Perplexity	Space system prompt	`Sub`
GitHub Copilot	`.github/copilot-instructions.md`	`Free` (per repo)
Other / any	Paste profile as message 1	`Free`

Safety¶

Audit your memory monthly. Open the memory pane and skim every line. Delete anything stale.
Treat memory like a public bio — assume an admin or auditor can read it on enterprise plans.
Never delegate "remember my password" or "remember this credential."
If your AI offers an "incognito / temporary chat" mode, use it for sensitive one-off questions.