ResearchFS turns papers, repos, lab notes, and experiment results into a memory graph your research agents can use. The next run does not start from a blank prompt. It starts from everything your lab has already learned.
A paper in a folder is just a file. A useful paper becomes a method, an implementation target, an experiment, and a result your research agents can retrieve later.
ResearchFS is not storage for PDFs. It is the working memory between your papers, codebase, experiments, and agents.
Humans search and inspect the graph. Research agents retrieve the methods, caveats, and prior results before touching code.
When an experiment works or fails, the lesson stays behind. Future research agents do not have to rediscover it.
The path is deliberately simple: prove one paper can create one useful code change, then turn that loop into team workflow, agent API, and enterprise research memory.
Ingest papers, extract methods, retrieve methods, run a code-edit research agent, log results, and compare baseline vs. ResearchFS.
Upload papers, ingest arXiv and GitHub, search methods, watch experiments, inspect provenance, and review what changed after each run.
Let any coding or research agent ask for relevant methods, prior outcomes, implementation notes, and lessons before it proposes a change.
Bring in private papers, internal repos, Slack, Notion, lab notes, experiment trackers, permissions, audit logs, and hardware-specific lessons.
Instead of pasting twenty papers into every prompt, a research agent asks ResearchFS what the lab already knows: which methods map to this code, where they came from, and what happened last time.
Teams will not buy “paper storage.” They will buy fewer repeated failures, lower context cost, fewer wasted GPU runs, and research knowledge that survives team churn.
Research agents retrieve the relevant method instead of stuffing entire papers into context.
People stop re-explaining the same papers, failures, and implementation details.
Bad ideas get filtered against prior results before they hit expensive training jobs.
When a researcher leaves, the method/result graph stays with the lab.
Agents propose stronger experiments because they can retrieve tested internal lessons.
In a local nanochat benchmark using zai-org-glm-5-2 via Venice, ResearchFS made the research-agent loop cheaper and faster. The baseline still won BPB, which is exactly the kind of lesson the memory graph should keep.
MLP SwiGLU during the GLM run.If your team is experimenting with autonomous research agents, the question is not whether they can read papers. It is whether they remember what those papers did in your codebase.