What to store in your AI memory, and what never to put there

Updated 30 October 2026. Advice that applies to any AI memory; Pryzm-specific behaviour checked in its code on that date.

Store in your AI memory what helps across conversations: your working preferences, your decisions and why, your projects' and clients' context, problems you already solved. Never put passwords, API keys, card numbers, ID numbers or health data in it. For everything else, ask: would I put this in a notebook shared with a trusted contractor?

This guide applies to ChatGPT's or Claude's built-in memory as much as to an external one. The specific examples are about Pryzm Memory, a persistent memory shared across several AIs over MCP.

Store: what makes your AIs useful

1. Your working preferences: “Short answers, British English”, “Strict TypeScript, no classes”. These pay off most: they help in every conversation.
2. Decisions and their reason: “We keep Stripe over Paddle, for EU VAT”. The reason stops an AI from suggesting the option you ruled out.
3. Your projects' and clients' context: goal, budget, deadline, contacts, constraints (“no purple”).
4. Solved problems: “421 error behind nginx: add the host to allowed_hosts”. The next AI won't start from scratch.
5. Your goals: “Ship the English version before the end of November”.
6. Things to follow up: “Chase the Atelier Lumen quote on Monday”.

Tip: one fact per memory, written to make sense on its own six months from now. “Atelier Lumen: delivery on 5 December”, not “it moved to the 5th”.

Never store

• Passwords, API keys, private keys, payment card numbers. Pryzm refuses them outright: nothing is saved, and the AI gets a message explaining why. They belong in a password manager.
• ID numbers, social security numbers, full IBANs. Pryzm's instructions ask AIs not to save government identifiers, but the server doesn't block them: it's on you.
• Health data, yours or anyone else's: it is sensitive data under the GDPR. Same rule: AIs are asked not to save it, the server doesn't filter it.
• Other people's personal data without a reason: keep the useful context (“Claire, contact at Atelier Lumen”), not their full details or private life.
• Trade secrets under NDA, unless your contract allows it.
• Passing moods: “I'm annoyed with this client” ages badly and will come back at the wrong moment.

Three habits for a clean memory

1. ## private mode. Start a message with ##: AIs connected to Pryzm are instructed not to save anything from it, and the server refuses any content starting with ##. Handy for thinking out loud.
2. “Ask me before saving” mode. In Pryzm, Settings → Saving memories, choose “Ask me before saving”. If you pasted the Pryzm instructions into your AIs, copy them again: they change with this setting.
3. Regular clean-up. In your memories list, delete what is stale. When a fact changes, tell your AI: it saves the new version, and the old one can be marked as replaced rather than overwritten.

What AIs can and cannot do

In Pryzm, a connected AI can search and add memories. It can neither edit nor delete them: there is no tool for that, by design. If an AI saved something you didn't want, delete it yourself in your Pryzm dashboard. A memory repeated word for word is reinforced, not duplicated.

What about security? Each Pryzm account is encrypted with its own key (AES-256-GCM), on servers in Germany. During a session, the server decrypts your memories in memory so it can search them: it is not end-to-end encryption. Hence this guide's rule: nothing you wouldn't trust to an online service. Details in what the word “encrypted” covers.

If something looks wrong

• An AI saved too much: delete the memory in your Pryzm dashboard.
• “Not saved: this content appears to contain…”: your text looks like a password, a key or a card number. Rephrase without the value.
• An AI brings back stale information: give it the new version, then delete the old one if you don't want it.

To check that your AIs really share what you tell them, run the 2-minute test. To start with a memory that's already full, see import your ChatGPT memory.