I’m not a developer, but I’ve gradually started using GitHub as persistent external memory for ChatGPT.
My use case is not primarily coding.
I use ChatGPT Web for long-form thinking: career decisions, business ideas, investing, product concepts, research, personal principles, and other evolving mental models.
The pattern I’ve ended up with is:
ChatGPT conversation
↓
exploration / reasoning
↓
durable conclusion
↓
Git repository
↓
future ChatGPT conversation reads it
The distinction that has become useful to me is:
Conversation = working memory
Repository = durable state
The conversation is intentionally disposable.
It can contain uncertainty, contradictions, dead ends, and temporary hypotheses.
Only conclusions that are worth preserving get written into Markdown files such as:
VISION.md
PRINCIPLES.md
DECISIONS.md
STATUS.md
research/
models/
projects/
evidence/
This has a few advantages that I did not initially expect.
First, the memory is model-independent.
If I change models or interfaces, the important state survives.
Second, it is inspectable and editable by the user.
Third, Git history makes changes in beliefs and decisions explicit rather than silently overwriting them.
For example:
I believe X
↓
new evidence Y
↓
I revise X
becomes part of the persistent history.
I know there is already a lot of work around AGENTS.md, repo-native context, and persistent instructions for coding agents.
What I find interesting is applying the same basic idea to ChatGPT Web and non-coding personal knowledge.
Instead of the AI product owning the long-term state, the user owns a portable external state that different models can read.
That makes me wonder whether ChatGPT could eventually support this more explicitly:
- designate a repository or directory as durable context
- define an entry-point file similar to
AGENTS.md - distinguish temporary conversation from user-approved durable state
- allow explicit write-back of conclusions after a conversation
Is anyone else using ChatGPT this way today?