I’ve been experimenting with Shared Projects and noticed a memory gap that becomes important in long-running collaboration.
Today, Shared Projects correctly isolate personal memory for privacy: once a project is shared, it uses project-only memory and does not access each member’s personal memories, custom instructions, or conversations outside the project.
I think that privacy boundary is the right default.
However, it creates another problem: a Shared Project can gradually learn what the project knows, but it has no clear way to maintain who each collaborator is within the context of that project.
Example
Imagine two people collaboratively running a creator business.
One person is the creator. She uses ChatGPT every day, and her personal ChatGPT gradually learns things such as:
-
her level of technical literacy
-
how experienced she is with agents
-
what kinds of explanations work well for her
-
recurring mistakes in her prompting
-
skills she has recently learned
-
how much complexity she can comfortably handle
The other person helps with strategy, marketing analysis, and reviewing how effectively she uses AI.
They create a Shared Project for the business so both can collaborate on:
-
market research
-
content strategy
-
competitor analysis
-
performance reviews
-
marketing experiments
-
agent workflows
-
prompt reviews
This is a very natural use case for Shared Projects.
But after the project is shared, the creator’s personal memory is intentionally unavailable.
That means the Shared Project may know a great deal about the business, while knowing surprisingly little about the person operating it.
For example, a project chat may recommend an advanced spreadsheet workflow without knowing that this collaborator is a beginner. Or it may continue explaining basic agent concepts even though that person has improved significantly over several months.
Two Different Types of Memory
I think Shared Projects need to distinguish at least two namespaces:
1. Project Knowledge
Shared facts and state:
-
strategy
-
decisions
-
experiments
-
datasets
-
brand guidelines
-
current hypotheses
-
historical results
2. Member Profile
Project-relevant context about an individual collaborator:
-
role in the project
-
expertise level
-
preferred communication style
-
responsibilities
-
current skills
-
recurring workflow difficulties
-
project-specific preferences
The second category should belong to a specific member, rather than being flattened into generic shared project memory.
For example:
-
Alice — creator, beginner with spreadsheets, intermediate with ChatGPT, prefers concrete step-by-step instructions
-
Bob — strategy/engineering collaborator, comfortable with data analysis and automation
Then ChatGPT could adapt the same project knowledge differently depending on who starts the chat.
Privacy Should Remain Explicit
I would not want Shared Projects to automatically expose a user’s global personal memory to collaborators.
Instead, this could be implemented with explicit controls.
Possible approaches:
-
Automatically build a project-specific member profile only from that member’s activity inside the project.
-
Let each member inspect and edit their own project profile.
-
Allow users to selectively import specific items from personal Memory into a Shared Project.
-
Clearly show which profile information is visible to other collaborators.
-
Allow some profile fields to influence ChatGPT privately without exposing their contents to other members, if technically and privacy-wise appropriate.
For example:
“Share my technical skill level with this project”
could be allowed without sharing unrelated personal memories.
Relationship to Transparent Project Memory
There is already an excellent community proposal for a visible and editable Project Memory Dashboard.
I think that solves one half of the problem: What does this project know?
This request addresses the other half:
Who are the people doing the work, and what project-relevant context should ChatGPT know about each of them?
For long-running collaboration, both matter.
A Possible MVP
A relatively small first version could add a “My profile in this project” section containing:
-
Project role
-
Relevant skills / expertise
-
Preferences
-
Working style
-
Optional notes
-
AI-suggested profile updates requiring user approval
Each collaborator would manage their own profile.
ChatGPT could then use:
Shared Project Knowledge + Current Member Profile
when answering a project chat.
Later versions could support selective import from global Memory.
Why This Matters
Shared Projects are moving ChatGPT from a personal assistant toward a collaborative workspace.
In a personal assistant, “understanding the user” is one of the most valuable effects of long-term memory.
That benefit currently disappears exactly when multiple people begin collaborating.
The privacy isolation makes sense, but I think there is room for a middle layer between:
“Expose my personal memory to the project”
and
“The project knows nothing about me outside the messages currently available inside it.”
A user-controlled, project-scoped member memory would preserve the privacy boundary while making Shared Projects much more capable as long-term collaborative workspaces.