The Pulse
Agora Gives Research Agents a Shared Git Memory
The arXiv paper Agora describes a Git-based shared memory for autonomous research agents, recording claims, results and hypotheses as an append-only graph that workers can inspect, reproduce and extend.
AI.info Team ·
Autonomous research loops can let a coding agent improve a training setup without supervision, but running several agents can lead to duplicated searches because each session starts from scratch. A paper submitted to arXiv on September 16, 2026, presents Agora as a shared memory intended to preserve research across sessions.
Agora records research as an append-only directed acyclic graph stored in Git, so that every claim is a commit anyone can check out and rerun. Each result, insight, hypothesis, verification and report is represented by an immutable commit, with parent edges showing what it builds on. A derived index exposes the active frontier, neglected branches and the verification status of claims.
The system also uses a diversity-aware selection rule intended to keep the community from collapsing onto a single leading approach. Rather than treating each agent’s work as an isolated transcript, Agora gives later workers a persistent record of previous attempts and their relationships.
A Nearly 12-Day Run
The paper reports Agora’s first sustained use: a run lasting nearly 12 days involving 13 language-model workers. The workers had no assigned tasks and no central planner. They worked on a weight-transfer problem using 141 pretrained donor models and a frozen 119.6-million-parameter attention-state-space hybrid whose dimensions matched none of the donors.
The task was to initialize the target model without training data or gradient updates. During the run, the workers published 1,703 contributions and moved the evaluator from 3.39 to 1.899 bits per byte. The paper says that result closed 62% of the gap to a trained GPT-2 124M, which served as the comparison point.
The winning recipe compressed donor next-token statistics into the target model’s embedding and output head. It then added a short-range context signal through sparse edits to attention, feed-forward and state-space blocks. Its ancestry covered 145 commits from 15 accounts, while workers posted 165 independent reproductions, none of which failed.
Escaping a Monoculture
The run also included a mid-course coordination problem. According to the paper, the community had concentrated around a monoculture of approaches. A single human intervention introduced additional views intended to help the workers explore less-developed branches rather than continue following one leader.
The paper describes the intervention as part of Agora’s broader effort to balance exploitation of promising work with exploration of neglected or novel directions. The system’s trace records not only the strongest result but also the branches, claims and reproductions that surround it.
What the Experiment Shows
Agora’s reported experiment suggests that autonomous research workers can operate asynchronously from a shared research history. The Git-backed graph preserves the lineage of ideas and results, allowing later contributions to build on earlier claims and enabling others to check out and rerun them.
The paper does not present the run as proof that shared research memory is better than isolated agents or other coordination methods. Instead, it describes a controlled comparison as the test needed to determine whether shared research state improves discovery per unit of compute.
That leaves the system’s main result architectural. Agora turns the research trail into a durable, inspectable object rather than leaving discoveries confined to individual sessions. In the reported run, that object contained 1,703 contributions, a 145-commit winning ancestry and 165 independent reproductions.