Skip to main content
SenseLab’s memory doesn’t stand still. The more your agents work, the more it knows about which memories are worth surfacing and how far they can be trusted. The ML layer is the part that does that learning. It comes with your account and runs on its own. There is nothing to switch on, no training infrastructure to stand up, and no data to ship anywhere — the models train inside SenseLab, on history your agents are already producing.

​
What it learns from

Outcome history, and nothing else. When an agent finishes a task, SenseLab records which memories were involved and how the work turned out. That pairing — what the agent knew, and what happened next — is the entire training signal. It accumulates quietly in the background while your agents do their jobs.
Failures teach as much as successes. A history of nothing but good news gives the models nothing to contrast, so memory that records only wins improves more slowly than memory that records both.

​
Sharper recall

Retrieval starts out blending a few generic signals — how closely a memory matches the question, how recent it is, how much confidence it carries — in fixed proportions. Sensible defaults, but the same ones every account begins with. They can’t know that in your account recency counts for more than confidence, or that the memories confirmed by real outcomes are the ones that consistently pay off. So SenseLab learns that from your history instead. Memories that were in play before work succeeded become examples of useful. Memories that preceded failures, or that nothing has ever validated, become examples of not useful. A model picks up the patterns that separate the two, weighing things like how confident a memory is, how old it is, how many outcomes have confirmed it, whether it’s a stated fact or a working hypothesis, and whether it’s genuine knowledge or just a stored working file. Once it has learned something, that judgement joins retrieval alongside the generic signals, and briefings and recalls start putting your proven memories first. Your agents don’t do anything differently — they ask the same questions and get better answers.

​
Better-calibrated trust

Every memory carries a confidence score: how much trust it has earned. Outcomes move it. Work that succeeds nudges the memories behind it up; work that fails pulls them down, and further than a single success lifts them, because trust should be easy to lose and slow to rebuild. How much each outcome should count, though, depends on what you do. A critical failure in a payments service doesn’t mean what one in a logging service means. So SenseLab measures it. It looks at what became of the memories behind each kind of outcome — whether they kept turning up in failures, or went on to support work that succeeded — and works out the adjustments that fit your own history. It also estimates how long memories in your account stay useful before they’re better treated as stale.
This is measurement, not silent surgery. SenseLab reports what your history suggests; the standard adjustments above keep applying unless someone deliberately changes them. Confidence scores never shift under you because a model decided they should.

​
Training better agents

This is the part that compounds. A decision trace is already shaped like training data: here is what the agent knew, here is what it decided, here is how it turned out. Context, action, result. That shape is exactly what fine-tuning wants, and producing it is normally the expensive part — someone has to sit down and label thousands of examples. Your agents generate it as a by-product of working, and the outcome is the label. Nothing has to be ranked by hand. From that history SenseLab can produce datasets in the shapes the training ecosystem expects: examples of decisions that worked, pairs contrasting a success against a failure so a model can learn preference, and memories scored by how their outcomes turned out so a model can learn to judge quality on its own. Those feed the standard tools — OpenAI’s fine-tuning API, Hugging Face, TRL — and what comes out the other side is a model that has internalised what your organisation already learned the hard way.

​
When it starts helping

Learning needs something to learn from, so a new account starts on the defaults: recall uses the standard signals, confidence uses the standard adjustments, and nothing is worse for it. The models begin contributing once there is enough history to be meaningful — on the order of twenty outcome-linked memories before ranking improves, and a handful of each kind of outcome before trust calibration will draw a conclusion. With agents recording outcomes as they work, that point arrives on its own within the first weeks of real use.

​
The loop

This is what separates memory from storage. Storage gives back what you put in. Memory that learns gives back more than you put in, because the work your agents do today is the training signal that makes them better tomorrow.