Controlled experiments on memory in AI systems, run with pre-registered decision rules and published in full, including the results that say no. And what those results tell us about the way people remember.
We gave progressive memory demotion its best possible chance: a perfect reader, tuned policies, a rule written before the data. It lost at every budget. Forgetting before you know the question destroys the history later questions need.
Each program has its own repository, pre-registration, and decision record. Each study ships code, a decision record and an essay, with a paper to follow.
Can memories demoted into compact traces, reinstated by context, beat keeping everything under a fixed context budget? A pre-registered oracle test says no.
What happens to a memory after it's made. Details will be published with the first result.
Pre-registration in progress
The decision rule is committed before the data exists. The commit history proves it.
Negative results are results. Our first finding is one.
Every bug we fixed, and whether it changed the answer, is on the record.
Code, configurations, and seeds for every figure, under version control.
Studies in progress stay private. When one is finished, its pre-registration, code, data and essay are published together, and the paper follows.
I tested whether an AI should let its memories fade the way ours do. In my test, it shouldn't, at least not before it knows the question.
Coming with the first Rewrite result.
Every repository records its git hash in every run. If a number is on this site, you can regenerate it.
Aftertrace is run by Shikhar Kumar, PhD, who trained in the cognitive science of learning and memory, with research on episodic memory, consolidation, and reconsolidation, and now builds production machine-learning systems. The lab returns to those questions with tools that did not exist then.
One email per finding: the result, the essay, and the code. No hype.
