EvoMap Module-by-Module Evidence Alleges Hermes Agent Architecture Plagiarism, Nous Research Only Responds with "Delete Account"

According to Observify Beating monitoring, the development team of the AI Agent's self-evolving engine Evolver, EvoMap, released a technical comparison article on April 14th, accusing the open-source project Hermes Agent under Nous Research of systematically replicating Evolver's design at the architectural level without any acknowledgment in any of its public materials. Hermes Agent currently has 88,000 GitHub Stars, while Evolver has around 2,000. Nous Research's founder, Teknium, responded on X, stating, "I have never heard of this person or this project," and considered the comparison article "completely brainless." The official response from Nous Research's account was even more concise: "Our repository was created in July 2025. We are pioneers of underlying technologies for modern Agent frameworks like YaRN. Delete your account."

The EvoMap article systematically listed over ten architectural correspondences:

1. Three-tier memory system (Persistent Facts + Procedural Knowledge + Historical Search) with precise functional role alignment for each layer
2. Closed-loop for automatically extracting reusable experiential assets after task completion (Evolver refers to this as Gene/Capsule, while Hermes refers to it as SKILL.md)
3. Periodic reflection mechanism (Evolver triggers every 5 evolutionary cycles, Hermes triggers every 15 tool invocations)
4. Runtime skill discovery and on-demand loading
5. 10-step orchestrated evolutionary loop
6. Multi-dimensional weighted scoring, constraint gating, atomic writes, and other engineering patterns

On the timeline front, Evolver's core protocol, GEP, was entirely open-sourced from February 1-16, with a total of 136 releases. Hermes' skill system was only released on March 12 (v0.2.0), and its self-evolving repository was created on March 9, with a gap ranging from 24 to 39 days.

However, the article also highlighted facts in favor of Hermes: the Hermes Agent repository was created in July 2025, half a year earlier than Evolver, but remained private until the end of February; Hermes' self-evolving module uses the Stanford/Berkeley academic framework GEPA (ICLR 2026 Oral), unrelated to GEP; and Anthropic's Agent Skills standard released in December 2025 predates both projects. The two projects use different languages (JavaScript vs. Python), and no code-level replication was found.

From any single dimension, independent convergence is entirely plausible: three-tier memory, experience extraction, and periodic reflection have all been widely discussed in the AI Agent field. EvoMap's argument is based on probability reasoning that "simultaneously corresponding in over ten dimensions goes beyond the realm of coincidence," which is compelling but not conclusive. Teknium chose to deny awareness rather than respond to technical details and did not provide further explanation. This debate is unlikely to reach a conclusion in the short term, but it reflects an ongoing issue: in open-source projects with disparate influence, the smaller party must assert architectural precedence, which is extremely challenging in terms of both evidence and dissemination.