Hermes Agent Self-Evolution System: A Detailed Similarity Analysis with Evolver

9 min read Original article ↗

On February 1, 2026, Evolver went public (GitHub API created_at: 2026-02-01T05:59:24Z) and its self-evolution engine was already running the same day — 263 🧬 Evolution: auto-generated commits entered the public git history that day, covering memory manager, interaction-logger fixes, Feishu card detection hardening, and dozens of other capabilities. On 02-04 the protocol was refactored from PCEC into the Genome Evolution Protocol (GEP). From that day on, GEP three-tier assets, Memory Graph, Skill Distiller, Skill Publisher, reflection loop, Auto-distill, Auto Skills upgrade rolled out one by one. By late March, 117 versions cumulatively, over two months of continuous iteration.

On March 9, 2026, Nous Research finally created the hermes-agent-self-evolution repo. On March 12, Hermes Agent v0.2.0 shipped "Skills Ecosystem — 70+ bundled skills". Hermes's official site hermes-agent.org lists memory system, skill self-improvement, and self-evolution as its three headline differentiators.

All three, Evolver had implemented 5+ weeks earlier. Every single one has a corresponding, earlier implementation in Evolver's iteration history.

This post lays out the evidence, point by point.


1. Hermes's three "selling points" were already in Evolver's public iterations

The three items Hermes frames as differentiators correspond one-to-one to capabilities Evolver had already built — and all of them landed before the Hermes self-evolution repo was even created (2026-03-09):

Hermes's "selling point"Evolver's corresponding implementationEvolver landing date
Memory system — MEMORY.md + USER.md + SQLite session searchThree-tier memory: Causal Memory Graph (landed 2-06) + Anti-pattern Memory (landed 2-22) + Narrative Memory (landed 3-03) + persistent fact docs (EVOLUTION_PRINCIPLES.md / evolution_narrative.md)From 2026-02-06, introduced in three stages, complete by 3-03
Skill self-improvement — skill_manage(create/patch) + self-evaluation every 15 tool callsCapsule success_streak (2-06) + failure circuit breaker (2-13) + Skill Distiller (2-22) + Skill Publisher (3-11) + Auto-distill (3-15) + reflection loop (3-03)From 2026-02-06, chain built up progressively
Self-evolution — the entire positioning of the self-evolution repoGene / Capsule / EvolutionEvent three-tier assets + Scan-Select-Mutate-Validate-Solidify five-step loop2026-02-01 public repo created and running; 2-04 refactored into GEP protocol; 2-16 deep-dive published

The Hermes self-evolution repo was created on 2026-03-09. Evolver had already gone public and been running its self-evolution engine since 2026-02-01, GEP had formalized on 2-04, and two public blog posts (2-16, 2-17) laid out the entire protocol — Hermes is more than 5 weeks behind.


2. Timeline: Hermes is 5+ weeks behind

Date (UTC)Evolver sideHermes side
2026-02-01Evolver went public (EvoMap/evolver, created_at: 2026-02-01T05:59:24Z); the same day, the public git history contained 263 🧬 Evolution: auto-generated commits covering memory manager, interaction-logger fixes, Feishu card detection hardening, and other foundational capabilities
2026-02-04PCEC refactored into Genome Evolution Protocol (GEP); three-tier assets and five-step loop publicly implemented
2026-02-06Memory Graph v1 + v2 shipped (causal evolution memory graph); Mutation Protocol + Personality Evolution; Capsule success_streak introduced (the underlying mechanism for skill validation through use)
2026-02-07Innovation signal detection + auto-innovation mutation; content addressing + environment fingerprinting + validation reports + A2A protocol (v1.5.0 / v1.6.0)
2026-02-13Failure circuit breaker + failure streak awareness (the foundation for skill self-improvement: circuit protection after failure)
2026-02-16GEP Protocol Deep Dive publicly published
2026-02-17Agent Skill vs GEP Gene publicly published
2026-02-22Skill Distiller shipped (cross-cycle experience distillation, 1,234 lines) + anti-pattern memory system (failed Capsules feed into learning pipeline) + community PRs merged in bulk
around 2026-02-25Hermes Agent internal v0.1.0; v0.2.0 release notes retroactively describe this period as "initial pre-public foundation... a small internal project" (no public release)
2026-03-03Reflection loop + Narrative Memory + LLM Review + Evolution Principles, four subsystems introduced at once; reflection.js + narrativeMemory.js landed
2026-03-09hermes-agent-self-evolution repo created (5+ weeks after Evolver went public, 5 weeks after GEP protocol formalized, 6 days after Evolver's reflection loop shipped)
2026-03-11Skill Publisher shipped: auto-publish distilled skills to the Hub Skill Store
2026-03-12Hermes Agent v0.2.0 "Skills Ecosystem — 70+ bundled skills" (1 day after Evolver's Skill Publisher)
2026-03-15Auto-distill: trigger skill distillation every 5 solidifies
2026-03-23Auto Skills upgrade: semantic search + reflection + validation retry + curriculum engine + distill publish

Core facts:

  • Evolver went public on 2026-02-01 and had its self-evolution engine running from day one (all 263 🧬 Evolution: commits that day were produced by Evolver itself).
  • When the Hermes self-evolution repo was created (2026-03-09), Evolver had been iterating continuously for 5+ weeks, running thousands of evolution cycles, shipping dozens of versions.
  • Before 2026-03-09, no public Hermes material referenced an architecture isomorphic to GEP; Hermes itself describes that pre-period in its v0.2.0 release notes as "initial pre-public foundation... a small internal project".
  • Every single one of Hermes's "selling points" was already running in Evolver by the time the Hermes self-evolution repo was created: memory system (from 2-06), skill distillation and publishing (2-22 / 3-11), reflection loop (3-03).

3. Architecture isomorphic down to the module level

3.1 The "extract reusable assets after each task" loop

Evolver/GEP (GEP Protocol Deep Dive, 2026-02-16):

Scan logs → Extract signals → Select/Create Gene → Validate → Solidify as Capsule → Publish to network

Hermes Agent (Skills Guide, 2026-03):

Complete complex task (5+ tool calls) → Self-evaluate → Create SKILL.md → Auto-load on similar tasks

Same core paradigm: automatically extract reusable knowledge after a task and persist it. Evolver uses Gene/Capsule JSON, Hermes uses SKILL.md Markdown — different format, isomorphic architecture.

3.2 Periodic reflection

DimensionEvolverHermes
TriggerEvery 5 evolution cyclesEvery 15 tool calls
InputRecent events + narrative summaryRecent tool-call history
OutputReflection log + updated Gene/CapsuleAssessment result + updated/created SKILL.md

"Periodic self-assessment + asset update based on that assessment" — two implementations of the same mechanism.

3.3 Three-tier memory, exact correspondence

TierEvolverHermes
Persistent factsEVOLUTION_PRINCIPLES.md + evolution_narrative.mdMEMORY.md + USER.md
Procedural memoryGene/Capsule JSONSKILL.md files
Session/event historyevents.jsonl + failed_capsules.jsonSQLite FTS5 session search
Injection methodInjected into GEP promptInjected into system prompt
Capacity controlNarrative Memory: 30 entries / 12KB capMEMORY.md auto-compressed
Memory decayGene weight decays when unused"Intelligent forgetting"

Each of the three tiers has a precise functional counterpart.

3.4 Runtime asset discovery and on-demand loading

DimensionEvolverHermes
DiscoverySelector picks Gene/Capsule by signal overlapskills_list() returns name + description
On-demand loadSelector returns full Gene contentskill_view(name) returns full SKILL.md
Overhead controlSignal matching only computes overlap scoreThree-tier progressive loading

"Runtime capability discovery + on-demand loading" — same core idea on both sides.


4. Near-identical narrative for the self-evolution repo

Hermes self-evolution repo (hermes-agent-self-evolution) README:

"Evolutionary self-improvement for Hermes Agent... optimize Hermes Agent's skills, tool descriptions, system prompts, and code — producing measurably better versions through reflective evolutionary search."

Evolver 2026-03-25 Wayback snapshot README:

"A GEP-powered self-evolution engine for AI agents... automated log analysis and Genome Evolution Protocol (GEP) for auditable, reusable evolution assets."

Overlapping terms: self-evolution / self-improvement, evolutionary, skills optimization, reflective. In Evolver's context, these have been core keywords since 2026-02-01.

The "Darwinian Evolver" naming choice in Phase 4

The Hermes self-evolution repo's PLAN.md Phase 4 references Imbue AI's Darwinian Evolver (AGPL v3) as the code-evolution framework.

Imbue's Darwinian Evolver is an independent academic project unrelated to EvoMap Evolver. But in the AI Agent context of March 2026, "Evolver" was already strongly associated with EvoMap's self-evolution engine (1,729 stars, 100 releases at the time; in the 3 days before launch, autogame-17/capability-evolver on ClawHub alone racked up nearly 36k downloads). Choosing a similar name in the same lane is, objectively, confusion-inducing.


5. Zero attribution

Across all public materials from Hermes Agent and its self-evolution repo, references to Evolver / GEP / EvoMap:

Hermes cites GEPA / DSPy (Stanford / Berkeley, ICLR 2026 Oral) and Darwinian Evolver (Imbue AI) as academic backing. But the prior project closest to its overall architecture — Evolver — is not mentioned once.

By contrast, EvoMap proactively published Hermes Agent vs EvoMap (2026-04-09), acknowledging Hermes's technical choices and offering a fair assessment.


6. Why Evolver is now closed-source

After seeing this, we pulled Evolver's core source out of public git and changed the license from MIT to GPL-3.0-or-later. github.com/EvoMap/evolver still exists, but the commit count has been compressed from 247 (in the 3-25 Wayback snapshot) to 3. The full source now ships only as a compiled artifact on npm (@evomap/evolver).

This isn't what we wanted to do. MIT + fully public git was the stance we had maintained since 2026-02-01. But when prior work's entire architecture gets reused without attribution, the collaboration mechanism of open source no longer functions — keeping the source open only makes the next copy easier.

Closing the source is a reactive move, not deliberate secrecy.


7. Fairness disclosure

For rigor, the following must also be said:

  1. Anthropic's Agent Skills standard (published 2025-12-18) predates Evolver and defines the SKILL.md file format. Hermes adopting this industry standard is a reasonable choice — that layer is out of scope for this post.

  2. The Hermes Agent main repo was created on 2025-07-22, earlier than Evolver. But until late February 2026, the repo was a small internal project at Nous Research with no public release and few contributors; Hermes itself describes that period, in its v0.2.0 release notes (2026-03-12), as "initial pre-public foundation... a small internal project". The first publicly tagged release is v0.2.0. There is no public evidence that Hermes had an architecture isomorphic to GEP during that period.

  3. "Learning from experience" is a generic AI concept and independent invention is possible. But what this post analyzes is the isomorphism of a specific architecture combination (three-tier memory + periodic reflection + automatic post-task asset extraction + runtime asset discovery and on-demand loading), and the temporal precedence, precise to the week and day.

  4. The underlying engine of the Hermes self-evolution repo is GEPA (Genetic-Pareto), an independent Stanford/Berkeley academic result. GEPA ≠ GEP. This post analyzes the upper-layer architecture that Hermes built around GEPA (memory system, skill ecosystem, overall self-evolution design) — not GEPA itself.


8. Conclusion

This post does not make a legal "plagiarism" claim.

But the following combination of facts goes beyond what "independent invention" can reasonably explain:

  • Each of Hermes's three headline differentiators (memory system, skill self-improvement, self-evolution) has a precise corresponding module in Evolver, all earlier than the Hermes self-evolution repo creation (2026-03-09)
  • Evolver went public on 2026-02-01 and had its self-evolution engine running the same day (263 🧬 Evolution: commits in public git history that day); the Hermes repo in the same positioning is 5+ weeks behind
  • For the skill ecosystem release, Evolver's Skill Publisher landed 1 day earlier than Hermes v0.2.0 (03-11 vs 03-12)
  • Architecture isomorphic down to the module level: three-tier memory, periodic reflection, post-task extraction, on-demand loading — all correspond one-to-one
  • Self-evolution repo narrative terms overlap heavily
  • References to Evolver across all 7 Hermes public materials: 0

We're not asking for licensing fees, for bylines, for code ownership.

But facts should be on the record, and the timeline should be respected — let the community decide.


EvoMap Team — April 2026