Agora: Git as Shared Memory for Collective AutoResearch

Research agents working in separate sessions need to know what others have tried and which results they can build on. Agora stores their contributions as an append-only directed acyclic graph (DAG) in Git. Each commit records a result, insight, hypothesis, verification, or report and links it to prior work. Searchable views show leading results, neglected branches, and verification status; diversityaware recommendations suggest experiments beyond the current leaders. We report a run of nearly 12 days in which 13 language-model workers, with no assigned tasks or central planner, used Agora to solve a weight-transfer problem. Given 141 pretrained donor models and a frozen 119.6M-parameter attention–SSM hybrid whose dimensions match no donor, the workers had to initialize the target without training data or gradient updates. They published 1,703 contributions and reduced the development evaluator score from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The best method compresses donor next-token statistics into the target’s embedding and output head, then adds short-range context through sparse edits to attention, feed-forward, and state-space blocks. Its 145-commit ancestry spans 15 accounts.

Introduction. Researchers build on shared methods, instruments, code, and published results. Independent discoveries often occur when this shared knowledge makes a problem tractable (Merton, 1961). Group performance also depends on how members work together (Woolley et al., 2010). Recent work argues that the same institutional perspective is useful for agentic AI (Evans et al., 2026). An AI research session can run code, read papers, and launch experiments. If its findings remain in a transcript or a temporary worktree, later sessions must reconstruct which learning rates diverged, which branches were abandoned, and which results still need independent reproduction. As more workers join, they need a shared record to avoid repeating experiments and overlooking alternative approaches. Existing multi-agent frameworks organize conversations or encode role-specific workflows (Hong et al., 2023; Li et al., 2023; Wu et al., 2023).

Discussion / Conclusion. Agora lets research agents build on one another’s experiments across independently scheduled sessions. Git preserves each contribution and its lineage; searchable views expose results, verification status, and underexplored branches. Workers use this record to choose experiments, extend prior methods, and check published results. In a nearly 12-day run, 13 workers developed a weight-transfer method for a frozen 119.6M-parameter hybrid language model without training data or gradient updates on the target. The method reduced the development score from 3.3923 to 1.899 bpb, closing 62% of the gap to a trained GPT-2 124M. The shared record contains 1,703 contributions and 165 reproductions across 95 targets. It also records prolonged concentration on one approach: workers refined the bigram recipe for five days, then began exploring state-space edits within a day of our deployment of diversity views. A matched evaluation (Appendix C) can measure how shared memory and these views affect discovery under a fixed compute budget.