A research agent can run an experiment and write down what happened. With many agents, the harder problem is deciding which record to read and which previous finding to trust.
Contributions become a graph
Agora stores results, hypotheses, insights, and verification reports as an append-only Git-backed directed acyclic graph. Parent links preserve provenance. Searchable views expose promising findings, neglected directions, and verification status; diversity-aware recommendations guide attention beyond the current leader.
In the reported weight-transfer study, 13 workers contribute over nearly 12 days. They publish 1,703 contributions and improve a frozen 119.6M-parameter target from 3.39 to 1.899 bits per byte without target training data or gradient updates. These figures describe the particular transfer task and evaluator, not a general scientific-discovery benchmark.
Memory needs structure
Consider two hypothetical entries: “method B scored well” and “method B reproduces entry A, changes its initialization, and was verified under configuration C.” The second is more useful because it separates a claim from its supporting execution and states what changed.
A flat conversation log can preserve the same words but make these relationships expensive to recover. A graph makes dependency explicit. It can also preserve unsuccessful branches, which may keep later workers from repeating an already understood failure.
What the study does and does not show
The paper records human interventions, including changes to diversity views and a worker prompt during the run. Most of the observed score reduction arrives early. Measuring the causal benefit of shared memory or diversity guidance therefore requires a matched comparison, rather than attributing every later improvement to one interface feature.
The broader design question
More memory is not necessarily better retrieval. A useful research workspace must help agents locate credible evidence, inspect the experiment behind it, and decide whether further exploration is worthwhile. That is an information-allocation problem as much as a storage problem.
This interpretation connects Agora to OSWorld-Pro: both make intermediate work inspectable, although one records research contributions and the other evaluates actions in a desktop trajectory.