As of June 5, 2026, ChatGPT memory is no longer just a small personalization feature. OpenAI's June 4 update on better memory for ChatGPT describes a more capable memory architecture that can synthesize useful context over time and show a reviewable memory summary. For a solo user, that may feel convenient. For a small team, it changes the operating question: what should become persistent context, and what should stay tied to one project, file or chat?
The useful version is easy to imagine. ChatGPT remembers that the team writes customer updates in plain language, that invoices use a specific naming pattern, that a weekly operations note always needs risks and blockers, or that product decisions should separate confirmed facts from open assumptions. Those memories can reduce repeated setup work.
The risky version is just as ordinary. ChatGPT remembers a temporary client preference after the contract changes. It carries one project's tone into another. It uses an old pricing assumption. It keeps personal or personnel context that was only mentioned during a messy draft. It treats a one-time exception as the normal way to work.
Start with a memory inventory, not a settings toggle. List the kinds of context the team would be comfortable reusing across future chats:
- Stable writing preferences.
- Formatting rules for recurring internal notes.
- Product names, team vocabulary and approved acronyms.
- Review habits, such as "flag unsupported claims".
- Non-sensitive workflow constraints, such as "always ask before deleting source data".
Then list the context that should not become general memory:
- Client-specific commercial details.
- HR, hiring, health or performance information.
- Passwords, tokens, access paths or private system details.
- Temporary campaign decisions.
- Draft legal, tax, security or compliance interpretations.
- Anything that belongs to one customer, deal or incident only.
OpenAI's Memory FAQ separates saved memories from reference chat history. Saved memories are the details ChatGPT can keep until they are deleted. Reference chat history lets ChatGPT use information from past chats when it seems useful, but the FAQ also notes that ChatGPT does not remember every detail from past chats. That distinction matters for teams: if a detail must be stable, put it in explicit instructions or a project file. If it is sensitive or temporary, avoid putting it into broad memory at all.
Use a monthly memory review for the people who rely on ChatGPT heavily. Ask ChatGPT what it remembers, check the memory settings, and remove anything stale, too personal or too broad. The FAQ says turning saved memory off does not delete memories that already exist, and fully removing something may require deleting the saved memory and the original chat where it appeared. Treat deletion as a cleanup process, not a single click someone will remember during a busy day.
Projects need their own rule. OpenAI's Projects in ChatGPT help page says project memory can keep context inside the same project, and that project-only memory must be chosen when starting a new project. For Business users, shared projects are set to project-only memory when shared. That is a helpful boundary, but it does not replace good setup. A project should still have a narrow purpose, clear files and a short instruction note explaining what context belongs there.
For client work, create separate projects by client or engagement. Do not use one general "client work" project if the team needs clean boundaries. For internal operations, separate recurring work from experiments: one project for weekly operations notes, another for testing a new sales-report format, and another for a temporary hiring plan. If the experiment becomes a process, move the stable instructions into the recurring project and leave the messy history behind.
Before a wider rollout, run five checks:
- Ask ChatGPT what it remembers about the work and compare it with the team's intended memory list.
- Test a new project with project-only memory and confirm unrelated context does not appear.
- Share a project with one teammate and check exactly what files, chats and members are visible.
- Put one stale assumption into a test conversation, then verify how it can be removed.
- Ask for an answer that should require current source material, and reject responses that rely only on remembered context.
The right default for a small team is selective memory, not maximum memory. Let ChatGPT remember stable work preferences that save time. Keep client facts, prices, policies and sensitive details in source documents, project files or systems of record. Memory should make the next chat faster; it should not become the hidden place where the team stores decisions it cannot audit.