Lessons that earn their place: how Memory decides what to keep
Not every result deserves to become a lesson. This note walks through the loop that turns an Autoresearcher's experience into something the next session can use, and the checks that stand between the two.
· 3 min · Mythein
The problem with remembering everything
Give a research session a goal and it produces a lot of experience: approaches that worked, approaches that looked like they worked, and a long tail of dead ends that were still worth running. A system that keeps all of it does not learn. It accumulates.
The continual research loop is built around a narrower question. Not what happened, but what, of everything that happened, should the next session start from?
Experience is useful. Evaluated experience is reusable.
That distinction is the whole design. Memory is not a transcript. It is the set of lessons that have passed a test.
Three stages, one direction
Every Autoresearcher session moves through the same three stages. They are the stations on the loop you see on the homepage.
| Stage | Input | Output |
|---|---|---|
| Explore | a goal, Skills, the current Memory | candidate approaches, run against a Science World |
| Evaluate | candidates and their checks | scores, with the evidence behind each one |
| Carry forward | scored candidates | proposed lessons, tested before they are kept |
Explore and Evaluate are where the work happens. Carry forward is where most systems get careless, so it is where we spend the most care.
What a proposed lesson looks like
A lesson is a short, testable claim about the world the session was working in. It names the condition it applies to, the action it recommends, and the evidence that produced it.
lesson:
world: catalyst-screen-v3
condition: "reaction temperature above the nominal window"
recommend: "re-check the baseline calibration before tuning the loading"
evidence:
runs: [r-2091, r-2094, r-2102]
delta: "+4.1% yield at the target selectivity, 3 of 3 runs"Nothing in that block is free text a future session has to interpret. The condition and the recommendation are what get tested.
The test a lesson has to pass
Before a lesson is retained it is replayed against earlier tasks from the same Science World. Three outcomes are possible:
- It helps. Sessions that start with the lesson reach the same or better scores in fewer runs. The lesson is kept.
- It does nothing. Scores are unchanged. The lesson is dropped; it was a coincidence dressed as a rule.
- It hurts somewhere. The lesson improves the task it came from but regresses another. It is dropped, and the regression is recorded, because a lesson that only works once is a trap for the next researcher.
The third case is the one that matters. A memory that can only grow is a memory that will eventually contradict itself.
What this loop does not do
It is worth being precise here, because the phrase self-improving invites the wrong picture.
THE LOOP UPDATES MEMORY, NOT WEIGHTS
The current improvement loop changes what a session knows when it starts. It does not change the model. The model families your team already uses are the same models before and after a lesson is kept. Super Turing provides the runtime; the Autoresearcher scoring loop evaluates candidates against a Science World. It does not train Super Turing.
That is a deliberate boundary. Weights are expensive to change and hard to audit. A lesson is a paragraph with a test attached. When one is wrong, you can read why.
A worked example
A session on a catalyst-screening world was asked to improve yield at the edge of the temperature window. It explored eleven candidates, four of which cleared the world's checks. Evaluation ranked them; the top two shared one move, re-checking the baseline calibration before touching the loading.
Carry forward proposed that move as a lesson. Replayed against six earlier tasks from the same world it helped in five and was neutral in one, with no regressions. It was kept. The next session on that world started with it, cleared the checks in three candidates instead of eleven, and proposed a further lesson of its own.
That is the shape of progress we are after. Not a bigger transcript. A shorter path.
If you run a lab, a company or a research organization and want to see the loop against your own research questions, write to us. The environments are the part we would like to build with you.