Feature Request: Selective & Shared Memory Across Projects
I would like to suggest a more flexible memory model for ChatGPT Projects: Selective and Shared Project Memory.
The current project memory system works well for keeping work organized, but for users who maintain several long-running projects, the available choices are still too coarse. In practice, some information should be shared across projects, while other information must remain strictly project-specific.
The use case
I work on several long-term fiction projects in ChatGPT.
For example, I may have:
-
Project A: a large science-fiction story
-
Project B: another unrelated story
-
Project C: smaller writing experiments
There is information that should ideally be shared between all of them:
-
how I prefer feedback to be given
-
my writing goals
-
how I work with drafts
-
what kind of critique is useful to me
-
general lessons I have learned while writing
-
recurring stylistic preferences
-
how ChatGPT and I have established our working process
But there is also information that absolutely should not be shared:
-
characters
-
fictional technology
-
timelines
-
locations
-
plot decisions
-
canon
-
discarded ideas
-
story-specific terminology
At the moment, the distinction is largely made at the project level. That creates a difficult tradeoff:
Either projects are isolated too strongly, or they risk sharing more context than I want them to.
What I would like instead is a hierarchy of memory.
Proposed model
1. Global User Memory
Information that applies everywhere.
For example:
“The user prefers direct criticism rather than automatic agreement.”
This is about the user, not any particular project.
2. Shared Memory Groups
Users could create optional memory groups and assign multiple projects to them.
For example:
Writing
-
Prototype
-
Rajapinta
-
Short Stories
These projects could share information about my writing process and preferences without sharing their story canon.
Other possible groups might be:
-
Work
-
Research
-
Home
-
Programming
-
University
-
Personal projects
A project could potentially belong to more than one group.
3. Project-Specific Memory
Information that belongs exclusively to one project.
For example:
“Richard Keller is the protagonist of Prototype.”
This should never become context for an unrelated story simply because both projects are categorized as writing projects.
4. Conversation-Level Memory Controls
Inside a project, individual conversations could have a simple option such as:
Use as project memory source: On / Off
This would be extremely useful.
Not every conversation inside a project represents current truth.
Some conversations contain:
-
brainstorming
-
rejected ideas
-
temporary experiments
-
alternative versions
-
exploratory discussions
-
outdated assumptions
Currently, users often have to remove or relocate conversations if they do not want them influencing project context.
It would be much cleaner if a conversation could remain inside the project for reference while being excluded from automatic memory retrieval.
5. Memory Scope for Individual Information
An even more powerful version could allow a memory item to have a scope:
-
Global
-
Memory Group
-
Project Only
-
Conversation Only
-
Do Not Remember
For example:
“I prefer character dialogue to remain imperfect and natural.”
Scope: Writing
while:
“Leonore Bell is approximately 45 years old.”
Scope: Prototype only
This would make ChatGPT memory much more predictable.
Why this matters
The real problem is not simply remembering more.
It is remembering the right information in the right context.
For long-term creative work, excessive memory can be almost as harmful as insufficient memory.
If context from different projects becomes mixed, ChatGPT may accidentally introduce:
-
characters from another story
-
incompatible worldbuilding
-
rejected concepts
-
outdated decisions
-
assumptions from an unrelated discussion
On the other hand, fully isolating projects means losing useful knowledge about how the user works.
Selective memory would solve both problems.
Example
Suppose I have two science-fiction projects.
One project involves experimental field physics.
The other involves a spacecraft transferring its consciousness into a human.
ChatGPT should know in both projects that:
I prefer plausible science fiction and want honest feedback about internal consistency.
But it should not assume that the field technology from the first story exists in the second one.
That distinction is currently difficult to express.
With shared memory groups, it becomes straightforward:
Writing Memory
→ writing preferences and collaboration style
Project A Memory
→ Project A canon
Project B Memory
→ Project B canon
Memory Source Transparency
A related improvement would be the ability to see why ChatGPT knows something.
For example, a memory detail could show:
Source: Global Memory
or:
Source: Writing Memory Group
or:
Source: Prototype → Character Notes
This would make memory errors much easier for users to diagnose and correct.
Instead of wondering why ChatGPT introduced a particular assumption, the user could immediately see where it came from.
Optional Advanced Feature: Memory Permissions
Projects could also have simple permissions such as:
Can read from:
-
Global Memory
-
Writing Memory
Can write to:
- Project Memory only
This would prevent a story-specific discussion from accidentally contaminating shared memory.
It would essentially give users lightweight control over information flow between projects.
Why I think this would improve Projects significantly
Projects are becoming more than folders for conversations.
For many users they are persistent workspaces.
Once people use them for months or years, memory architecture becomes extremely important.
A hierarchical model such as:
User → Memory Group → Project → Conversation
would make ChatGPT much better suited for serious long-term work while still keeping the experience understandable.
The key idea is simple:
Users should be able to choose which context is shared, rather than having to choose between remembering everything and isolating everything.
I believe this would be useful far beyond fiction writing—for research, software development, professional work, education, business projects, and any situation where several related but distinct projects are maintained over time.