I’ve been working on a deceptively simple problem:
Can I preserve useful AI working context outside the conversation and recover it in a fresh session?
Not by copying the entire chat. Not by relying on the model’s memory. And not by telling the new model what I expect it to say.
I wanted to see whether a small, structured archive could stand on its own.
So I ran a deliberately narrow experiment.
The test
I created a minimal KISS archive containing the state of one topic and a designated artifact.
Then I gave that package to a fresh AI session.
The fresh session’s job wasn’t to continue the project or improve anything. It had a much smaller job:
Recover what was actually archived.
That distinction matters.
If the new session started filling gaps, advancing the project, or reconstructing things that weren’t in the archive, I wouldn’t know whether I had demonstrated portability or simply watched a capable language model improvise.
So the boundary was intentionally strict.
What happened
The fresh session successfully recovered the archived topic state and designated artifact.
Just as importantly, it respected what the archive didn’t contain.
Unknown or not-started project states remained unknown. The session didn’t silently redesign the project, advance its state, or manufacture missing history.
For the thing I actually tested, that’s a PASS.
What this does NOT prove
This is the part I care about most.
I have not demonstrated that an entire AI working relationship can be moved between models.
I have not demonstrated behavioral continuity.
I have not demonstrated that a fresh session will make the same decisions as the original session.
And I haven’t yet demonstrated that interrupted work can be successfully resumed from the recovered state.
The result is much narrower:
A minimal external archive was sufficient for a fresh session to accurately recover a bounded piece of archived working state and its designated artifact.
That’s useful.
But it’s only the first rung.
The next question
The next experiment gets harder.
Instead of asking a fresh session merely to describe recovered state, I want to interrupt real work at a point where the correct next action depends on what happened earlier.
Then branch from that same interruption point:
- the original session continues normally;
- a fresh session continues using only the archived state;
- another fresh session continues without the archive.
Now the question changes from:
“Can it recover the state?”
to:
“Does the recovered state actually govern what it does next?”
And after that:
“Can it successfully continue the work?”
Those are different claims, and I don’t want to collapse them into one.
Why I’m publishing this
A lot of my work with AI currently happens where nobody else can inspect it: conversations, local experiments, artifacts, failed tests, protocol changes, and notes.
I’m going to start exposing more of that work.
Not as polished claims that I’ve figured everything out.
As receipts.
What did I test?
What actually happened?
What evidence do I have?
What can I reasonably conclude?
What remains unproven?
This experiment is a small one.
That’s intentional.
The goal isn’t to make the experiment sound bigger.
The goal is to make the evidence easier to inspect.