The research questions that matter most are rarely answered in one sitting. Reaching a target, comparing a set of designs or closing an open question takes a campaign: many runs, several people, more than one tool and days or weeks of work. The value sits in the sequence, in what was tried, what came back and why the next step was chosen.
AI agents now help with campaigns like this, and most of them work one session at a time. This post looks at what that costs a long piece of research, and at how MS Duplex keeps the work in one piece through pauses, hand-offs and model changes.
Where long research loses its thread
In a typical setup, the plan for a campaign lives in a chat window, a notes file and the head of the person running it. Each of those has a weak point, and a long campaign finds every one of them.
- A session ends, and the plan, the progress and the reasons behind each step end with it. The next session starts by rebuilding them.
- A researcher hands the work to a colleague, and the reasons behind earlier choices stay with the person who made them.
- The team moves to a new model, and whatever the old setup had been told has to be told again.
- A new result replaces an earlier estimate, and a pasted transcript brings the old value back as if it were current.
- An unexpected result should change the route, and a fixed script carries on as written.
Each of these is small on its own. Over a campaign they add up to repeated runs, repeated questions and experts spending the start of every session explaining the project to the AI again.
A campaign kept as one job
MS Duplex keeps the campaign itself, beside the agent and the model. A researcher gives it the goal rather than a script, along with the constraints that matter, such as the budget and the steps that need a person’s approval. MS Duplex lays out the route from where the work stands to the goal, with a checkpoint at each stage your experts can review.
As the agent works through your tools, the result and the evidence at each checkpoint are recorded. Progress is on record at every stage, and so are the reasons for each choice. A pause, a hand-off or a new session costs a moment rather than the work.
A plan is useful only while it fits the facts. When a result changes the picture, MS Duplex plans again from that point, keeps what still holds and records why the route changed. A script fixes every step in advance. A plan in MS Duplex follows the evidence as it arrives.
Two other capabilities keep the route efficient. Before a costly step, the options are played forward against what your team has already confirmed, so the budget goes to the runs most likely to move the campaign on. And what each checkpoint confirms is learned at once, so the rest of the campaign, and the next campaign, starts from it.


What a new session starts from
Research memory is the other half of a long campaign. When a new session opens, on the same agent or another one you authorize, and on any model, it starts from a compact picture of where the work stands: the goal, the decisions made so far, the methods your team approved and the progress on each step. Every part of that picture keeps its source.
The agent picks up from there. Nobody replays a transcript or explains the project again, and the agent does not ask questions the team answered earlier. Older detail stays on record and out of the way, so the session works from what matters for the next decision.
When a fact changes partway through
Long campaigns produce facts that move. A measurement replaces an earlier estimate, or a reviewer revises an assumption the plan relied on. MS Duplex moves the memory on with them: the newest confirmed value is the one your agent uses from the next request, and the history of what changed, when and why stays on record.
- 0% vs 57%stale answers after facts changed: MS Duplex beside standard retrieval, in MatterSpace’s controlled tests
That comparison comes from MatterSpace’s controlled tests. Standard retrieval finds passages that look relevant, so the old and new versions of a fact can both come back. In a long campaign, where estimates are replaced as the work goes on, one stale value can send a whole stage in the wrong direction.
One campaign across sessions, people and models
Say a team’s agent is working toward a target through a series of runs, with one expensive confirmation step planned near the end. Several people touch the work, and more than one model. Here is how it could hold together:
- The lead sets the goalThe agent receives the target, the budget of runs and the steps that need approval, including the confirmation step. MS Duplex lays out the route in stages, each ending at a checkpoint.
- A checkpoint changes an estimateResults from the first stage replace an estimate the plan relied on. The memory moves to the new value, the history shows when and why it changed, and the route is planned again from that checkpoint.
- The work pausesThe team turns to other priorities for a while. When the work resumes, the session opens on the current state, focused on the next decision, with every earlier step on record.
- A colleague takes overA colleague opens a new session through a different agent, on a different model. It starts from the last checkpoint with the goal, the decisions and the new value in view, and asks nothing the team has already answered.
- The costly step waitsIn a governed environment, the confirmation run waits at its checkpoint for a person’s approval. The approval goes on record, with who gave it.
- The lead reviews the campaignThe lead opens the campaign record and sees every checkpoint, every change of route and what each step rested on.


What research leaders get from it
For the people responsible for a campaign, the gain is visibility as much as continuity. A campaign that keeps its thread is also one you can inspect at any stage.
- One record per campaign. Work that spans people, tools and sessions stays one campaign with one plan.
- Checkpoints to review at any time. Your experts can read the plan, the progress and the evidence at any stage, correct a step or change the goal.
- Hand-offs that keep their reasons. A colleague, or a colleague’s agent, takes over with every decision and its reason in view.
- Decisions another researcher can rebuild. Every step keeps its source, so the route to a result can be checked from the record.
- Experience that outlasts the campaign. What was confirmed along the way stays in a store your organization owns, ready for the next campaign.
Common questions
What happens when a session ends in the middle of a campaign?
The campaign carries on from its last checkpoint. The next session, on the same agent or another, opens with the goal, the decisions made, the evidence gathered and the progress so far.
Can a colleague on a different model pick up the work?
Yes. The plan and the memory live in MS Duplex, beside the model, so the agents you authorize start from the same state of the work on whichever model they use. Your rules decide which layer of memory each person and agent can reach.
Can our experts change the plan partway through?
Yes. Every checkpoint can be reviewed at any time, and an expert can correct a step or change the goal. MS Duplex plans again from that point and records why the route changed.
How is this different from saving the chat history?
A saved transcript holds the words of past sessions and leaves the agent to work out which parts still hold. MS Duplex keeps the current state of the work, with confirmed results, decisions and approved methods, each with its source, and moves on when a newer confirmed value replaces an older one.


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.









