A new preprint on arXiv introduces Epi, an autonomous research agent built to sustain investigations across model sessions. Unlike typical agents that reset after each interaction, Epi keeps a persistent research history and uses hierarchical memory to organize what it has learned. This allows it to pick up where it left off rather than starting from scratch.
The system is also designed to be self-improving. By accumulating experience from past research tasks, Epi can evolve its own methods over time, potentially making it more effective at tackling complex problems. The authors frame this as a move toward persistent, autonomous agents that don't just execute single tasks but carry out long-running scientific investigations.
Because this is a single source, the claims are presented as the authors' own description of their system. No independent evaluation or comparison to other agents is included in the abstract, so the practical performance of Epi remains to be seen.