More AI tools now say they learn. For a research team, that word covers very different things: a notes file the assistant rereads, a search over past documents, a model retrained on the team’s data, or a system that changes its behavior as results come in. Each has its uses. They do not give a team the same thing, and a short demo rarely shows the difference.
This guide is for the research or engineering leader who has to decide. It sets out six questions to put to any AI that claims to learn, with what a good answer looks like for each. Then it answers them for MS Duplex, so you can hold us to the same standard as everyone else.
Why “it learns” deserves a closer look
When an AI learns from your work, it starts to act on things nobody typed into the prompt that day. That is the point. It is also why a learning system calls for harder questions than a static one. A model that behaves the same on every request can be reviewed once. A system that changes with use has to stay reviewable as it changes.
The common approaches each answer part of this. Instruction files and notes keep what someone wrote down, and stay as current as the last edit. Retrieval brings back passages that look relevant to the case at hand. Fine-tuning changes what a model knows across the board, through a prepared dataset and a training run, and it fits well when that is the goal. Before a system starts shaping decisions, your team should know which of these it is buying, or what it adds beside them.
The cost of a poor choice shows up later, in the work. An agent repeats a failed step because a lesson did not stick. A stale value survives in a recommendation because nobody could find where it came from. A team stays on an older model because moving would lose what it taught the old one. The questions below surface those risks before a pilot, while they are cheap to discuss.
Six questions to ask any AI that claims to learn
Put these to every system on your list. Ask for a demonstration on a case your team chooses, and for the record behind each answer.
- What does it learn from?A good answer names the inputs: confirmed results, measurements, expert corrections with their reasons, approved decisions. Be cautious if the answer is “everything the user types.” In research, a passing remark and a confirmed measurement deserve different weight, and your team should decide which outcomes count.
- How can you see why its behavior changed?Ask to see one change in behavior and the evidence for it. A good answer shows what the agent now does differently, since when and which outcomes justified it. If the only explanation is that the system “has improved,” you cannot defend that improvement to anyone else.
- How do you correct or remove one lesson?Mistakes enter every record: a faulty measurement, a correction that held only under narrow conditions. Ask what it takes to fix one lesson, whether anything else changes when you do, and whether a removed lesson can be restored. A good answer involves no training run.
- What happens to the experience when you change models?Research teams change models often. Ask whether the experience is tied to one model’s weights or kept beside the model, and whether a new model starts with what your team has built or starts again.
- Who owns what it learns, and who can see it?Ask where the experience is kept, who owns it and whether it can be exported. Ask how sharing works inside your organization: whether one researcher’s habits can shape everyone’s agent, and whether anything learned from your work reaches another company.
- How will you judge it?A good vendor proposes a test on your own work, with a measure your experts agree before anything runs and results scored without knowing which setup produced them. Figures from someone else’s tests show what is possible. Your own test shows what it does for your team.
How MS Duplex answers on learning and evidence
MS Duplex is a learning layer that plugs into the AI a research team already uses: Claude, Codex, Cursor, any agent that speaks MCP, and a team’s own agents through the API and Python SDK. It learns beside the model, with its weights left as they are. Here is how it answers the first two questions.
MS Duplex learns from the outcomes and expert corrections your team chooses to capture: a test that passes, a measurement that comes back, a conclusion an expert approves, a correction with the reason behind it. Each lesson is kept with its evidence and the conditions it was confirmed under, and it applies from the next request. Where a new case falls outside those conditions, MS Duplex flags the difference instead of applying the lesson as if nothing had changed.
When a fact changes, the newest confirmed value applies and the history of the change is kept. That matters for any team whose reference values, calibrations or operating limits move as the research does.
On the second question, every change in an agent’s behavior is on record. What worked becomes the first thing the agent tries. What failed under certain conditions stops being proposed under those conditions. Each change is tied to the outcomes that justified it, so when a lead asks why the agent behaves better today than it did last month, the answer is in the record with its sources.


MS Duplex also keeps the line between what is known and what is assumed. It records which action led to which outcome, under which conditions, and tells your researchers when a link has not yet been established. Where the evidence is thin or a case is unlike anything seen before, it hands the decision to your model or flags it for a researcher, with the reason.
How MS Duplex answers on control, ownership and models
Correcting and removing a lesson
Each piece of experience shows where it came from, in one signed history your team can review and export. Correct a lesson and only that lesson changes. Remove it and the recommendations that depended on it go with it, while everything else stays. A removed lesson can be restored if your team changes its mind. None of this needs a training run, because nothing was trained into the model in the first place.
Your rules sit in the same place. MS Duplex keeps rules and preferences apart, and both stay visible in the reasoning behind a step. In governed environments, which are supported setups with hooks, a step that breaks a rule is stopped before it runs, and the steps you select, such as an expensive simulation or a change to shared data, wait for a person to approve them. A new rule applies from the next request.


Ownership and model freedom
The experience your team builds lives in MS Duplex, in a store your organization owns, whichever model or vendor you use. Over time it holds:
- Confirmed results, each with the conditions it holds under
- Expert corrections, each with its reason
- Decisions, with the evidence behind them
- Verified methods your agents reuse and adapt
- Plans, checkpoints and progress on long campaigns
Because that experience sits beside the model, a new model starts with all of it from its first request. Your team can use one model for hard analyses and another for recurring work on the same project, with both drawing on the same experience.
Inside your organization, each researcher, each team and the whole organization has its own layer. A scientist’s or engineer’s working habits stay personal, what a team confirms is shared with that team, and what the organization adopts as standard applies everywhere. Your rules decide what moves between layers, people and agents reach only what their scope allows, and none of it is shared with another company. Where MS Duplex runs and how your data is handled are agreed with your security and data owners from the first conversation.
Judging it on your own work
The sixth question is the one that decides. MS Duplex is evaluated through a pilot on your own work, beside your current setup, and judged by your experts.
Choose recurring workflows where decisions repeat and a wrong or unnecessary run is expensive. Agree the measure with your experts before anything runs: the cost of each accepted, reproducible result, and the expert time it takes. Run the same model, tools, data and budget with and without MS Duplex. Then have your experts score the results without knowing which setup produced them.
Hold every system you consider to the same test. An AI that improves your work should show it on your work, under a measure you set, scored by people who did not know which answer came from where.
Common questions
Does MS Duplex change the model we use?
It leaves the model’s weights as they are. MS Duplex learns beside the model, which is why a lesson applies from the next request, why one lesson can be corrected on its own and why the experience comes along when you change models.
Does MS Duplex replace fine-tuning?
Fine-tuning keeps its place for changing what a model knows across the board. MS Duplex works beside a fine-tuned model as well as a general one, and makes either of them learn from your team’s confirmed results.
Can we audit what the AI learned?
Yes. MS Duplex keeps a signed history of what was learned, used and approved. Your team can review and export it, rebuild a decision from it, and inspect, correct, remove or restore any lesson.
Who decides what the AI learns from?
Your team does. MS Duplex learns from the outcomes and corrections your team chooses to capture, and your scientists and engineers decide what is adopted. In governed environments, the steps you select wait for a person before they run.


Built by MatterSpace
MatterSpace builds MS Duplex for scientists, researchers and engineers in every field, so the AI they already use learns from every result they confirm.









