GitHub Trending Weekly 2026-04-13: Hermes Agent Hits 65K Stars, Persona Distillation Wave, and On-Device AI Infrastructure Taking Shape

Data period: April 5–13, 2026 (rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia

TL;DR: Hermes Agent added 32,572 stars this week, leading the growth chart by a wide margin. The persona distillation wave sparked by nuwa-skill kept rolling — more than half of the newly emerged repos this week are variations of "distill someone's thinking style into a Skill." The archival of nvim-treesitter triggered a significant Neovim community controversy, becoming the most-discussed open source event on HN this week (176 points). Google's edge AI push (gallery 🔁, LiteRT-LM 🔁) continued as a monthly staple alongside NVIDIA PersonaPlex, together sketching out the emerging infrastructure landscape for on-device AI.

📈 Fastest Growing — Top 11 Weekly Star Gainers

Source: github.com/trending?since=weekly 🔁 = also appearing on monthly trending (sustained momentum signal)

#	Project	+Stars/week	Total Stars	Language	Created
1	NousResearch/hermes-agent 🔁	+32,572	65,964	Python	2025-07
2	microsoft/markitdown	+8,202	104,500	Python	2024-11
3	HKUDS/DeepTutor	+5,560	17,213	Python	2025-12
4	multica-ai/multica	+5,362	9,286	TypeScript	2026-01
5	forrestchang/andrej-karpathy-skills	+4,969	16,507	—	2026-01
6	google-ai-edge/gallery 🔁	+4,369	20,660	Kotlin	2025-03
7	NVIDIA/personaplex	+2,905	9,079	Python	2026-01
8	TheCraigHewitt/seomachine	+2,698	5,783	Python	2025-10
9	coleam00/Archon	+2,410	16,998	TypeScript	2025-02
10	google-ai-edge/LiteRT-LM 🔁	+2,196	3,536	C++	2025-04
11	nvim-treesitter/nvim-treesitter	+154	13,647	Tree-sitter Query	2020-04

🆕 Top New Repos — Top 15 Newly Born This Week

Source: GitHub Search API (created:2026-04-05..2026-04-13, sorted by total stars)

#	Project	Total Stars	Language	Created
1	MemPalace/mempalace	43,367	Python	2026-04-05
2	alchaincyf/nuwa-skill	8,453	Python	2026-04-05
3	garrytan/gbrain	6,210	TypeScript	2026-04-05
4	alchaincyf/zhangxuefeng-skill	5,269	—	2026-04-05
5	farzaa/clicky	3,936	Swift	2026-04-07
6	xixu-me/awesome-persona-distill-skills	3,404	JavaScript	2026-04-06
7	LaurieWired/tailslayer	2,091	C++	2026-04-05
8	alchaincyf/hermes-agent-orange-book	2,088	—	2026-04-08
9	KKKKhazix/khazix-skills	1,709	Python	2026-04-06
10	hotcoffeeshake/tong-jincheng-skill	1,590	—	2026-04-05
11	yizhiyanhua-ai/fireworks-tech-graph	1,530	Python	2026-04-10
12	fikrikarim/parlor	1,417	HTML	2026-04-05
13	mattmireles/gemma-tuner-multimodal	1,229	Python	2026-04-07
14	nashsu/llm_wiki	907	TypeScript	2026-04-08
15	phuryn/claude-usage	878	Python	2026-04-07

Weekly Spotlight — Fastest Growing Top 11

📈 #1 — NousResearch/hermes-agent｜The Open-Source AI Agent That Evolves Itself

"The agent that grows with you"

+32,572 ★ this week｜65,964 total｜Python｜MIT｜🔁 Monthly sustained momentum

Hermes Agent dominated this week with 32K new stars, making it one of the most enduringly popular AI agent frameworks on GitHub Trending over the past several months.

Its core proposition is closed-loop self-evolution: the agent doesn't just run tasks — it extracts skills from every conversation, automatically refines those skills, and builds a cross-session user memory model. The "agent that keeps growing" pitch isn't marketing fluff: it genuinely implements DSPy + GEPA (Genetic Evolution Prompt Architecture, ICLR 2026 Oral) for self-improvement.

This week's explosion was triggered by v0.8.0 (released April 8): 209 merged PRs, adding Browser Use integration, remote backend support (runs on a $5 VPS or serverless environment), and worktree parallelism. It deploys across Telegram, Discord, Slack, WhatsApp, Signal, and CLI — and doesn't lock you into any single LLM (supports Nous Portal, OpenRouter 200+ models, OpenAI, and more).

Hermes Agent was born in July 2025, grew quietly through February, and has now become the go-to host platform for Persona Skills. That growth trajectory is worth watching.

📈 #2 — microsoft/markitdown｜The Swiss Army Knife for Converting Docs to Markdown

"Python tool for converting files and office documents to Markdown."

+8,202 ★ this week｜104,500 total｜Python｜MIT

markitdown is Microsoft's document conversion tool that converts Office files, PDFs, HTML, images, and more into Markdown — a standard preprocessing step in RAG pipelines and AI content workflows.

It re-entered the top-2 this week, likely driven by renewed community focus on AI document parsing. HN saw related projects emerge in parallel (a Go port of markitdown, LiteParse, and similar Show HNs), signaling this problem space is still actively evolving. Crossing 100K stars confirms it's become the practical standard in this niche.

📈 #3 — HKUDS/DeepTutor｜An Agent-Native Personalized Learning Assistant

"DeepTutor: Agent-Native Personalized Learning Assistant"

+5,560 ★ this week｜17,213 total｜Python｜Apache-2.0

DeepTutor comes from the Hong Kong University Data Intelligence Lab (HKUDS) and positions itself as a deeper AI tutor than ChatGPT for learning workflows.

What sets it apart isn't just "answering questions" — it's the unified workspace design: Chat, Deep Solve, quiz generation, deep research, and math animation all share the same context. You complete a full learning cycle without switching tools. TutorBot also has persistent memory — it remembers where you got stuck, sets reminders, and can learn new skills, making it feel more like a stateful AI companion than a stateless Q&A tool.

If you're building or using an AI-assisted learning workflow, DeepTutor's architecture is worth studying.

📈 #4 — multica-ai/multica｜Managing Coding Agents Like Real Teammates

"Turn coding agents into real teammates — assign tasks, track progress, compound skills."

+5,362 ★ this week｜9,286 total｜TypeScript

Multica solves a real pain point: when you're running multiple Claude Code, Codex, or OpenClaw sessions at once, how do you coordinate them? Most people's current answer is "manually switch terminals." Multica offers a structured multi-agent collaboration platform instead.

Agents in Multica aren't passive tools — they're "teammates" with their own profiles. They claim tasks, report progress, create issues when blocked, and update their own status. Skills are shared across agents: define once, use across the whole team. The key trust property: all code runs on your local machine or your own cloud. Multica's servers only coordinate task state — they never touch your code.

Supports Docker Compose, single binary, and Kubernetes deployment. HN users observed that Multica's core value is closer to "GitHub Issues + Jira for agents" than a simple task queue.

📈 #5 — forrestchang/andrej-karpathy-skills｜Karpathy's AI Coding Principles as a Claude Code Skill

+4,969 ★ this week｜16,507 total

This repo contains a CLAUDE.md — Andrej Karpathy's (former OpenAI research scientist, Tesla AI director) observations on common AI coding pitfalls, curated by forrestchang into a directly usable Claude Code Skill format.

It's not a tool — it's distilled prompt engineering knowledge. Last week nuwa-skill went viral with a similar approach, and this Karpathy Skill reinforces the same pattern: in the Skills ecosystem, "distilling a knowledge worker's decision framework into executable instructions" is becoming a new form of knowledge distribution.

📈 #6 — google-ai-edge/gallery｜The Official Showcase for Local GenAI on Android

"A gallery that showcases on-device ML/GenAI use cases and allows people to try and use models locally."

+4,369 ★ this week｜20,660 total｜Kotlin｜Apache-2.0｜🔁 Monthly sustained momentum

Google AI Edge Gallery is an Android app (Android 10+ required) letting users experience various on-device ML/GenAI use cases directly on their phones — image classification, object detection, speech recognition, and the latest Gemma 4 local inference.

It has appeared on the monthly trending chart for consecutive weeks (🔁), signaling this isn't a one-off spike but a deliberate strategic push. Paired with LiteRT-LM (#10, also 🔁) and Google Developers Blog's concurrent article on "Gemma 4 Edge Agent Skills," it's clear Google is building its entire edge AI ecosystem around Gemma 4 as the core model.

📈 #7 — NVIDIA/personaplex｜Full-Duplex Voice AI That Listens and Speaks Simultaneously

"PersonaPlex code."

+2,905 ★ this week｜9,079 total｜Python｜MIT

PersonaPlex is NVIDIA's 7B-parameter full-duplex conversational model, open-sourced in January. Its technical core compresses the entire voice pipeline into a single Transformer — no ASR → LLM → TTS chain required. It does voice-in to voice-out directly, with speaker turn latency of 0.07 seconds (Gemini Live takes 1.3 seconds).

PersonaPlex lets you control the AI's "persona" via text role prompts and voice style settings, and naturally handles interruptions and backchanneling ("uh-huh," "I see," etc.). It's based on the Moshi architecture, weights are available on Hugging Face under a commercial-friendly license. This week's resurgence likely tracks with the rising overall interest in on-device voice AI (alongside gallery and LiteRT-LM).

"A specialized Claude Code workspace for creating long-form, SEO-optimized blog content."

+2,698 ★ this week｜5,783 total｜Python｜MIT

seomachine is a Claude Code workspace designed specifically for SEO content production, providing an end-to-end workflow covering research, writing, analysis, and optimization. Its existence signals something important: the user base for AI coding tools is no longer just engineers — content marketers and SEO practitioners are now adopting Claude Code as their own working environment.

This trend is worth watching. As Claude Code's Skills system matures, non-engineering AI workspaces (SEO, design, legal) could be the next wave of open-source repo explosions.

📈 #9 — coleam00/Archon｜YAML-Defined Workflows for AI Coding

"The first open-source harness builder for AI coding. Make AI coding deterministic and repeatable."

+2,410 ★ this week｜16,998 total｜TypeScript｜MIT

Archon's positioning is clear by analogy: Dockerfile is to infrastructure as GitHub Actions is to CI/CD as Archon is to AI coding workflows. You define your development process in YAML (plan → implement → validate → code review → PR creation), and Archon ensures every agent run is deterministic and repeatable.

It ships with 17 preset workflows, each running in an isolated git worktree to prevent cross-task contamination. Worth noting: the original Python version (task management + RAG) has been archived; the current version is a complete rewrite as a workflow engine. If you're trying to make AI coding produce consistent results rather than different outputs every time, Archon's design is worth a deep look.

📈 #10 — google-ai-edge/LiteRT-LM｜Google's Edge LLM Inference Engine

+2,196 ★ this week｜3,536 total｜C++｜Apache-2.0｜🔁 Monthly sustained momentum

LiteRT-LM is Google's C++ edge device LLM inference engine for running local language models on Android and iOS. It keeps appearing on the monthly trending chart as the underlying execution engine for both gallery and parlor.

Google Developers Blog published a concurrent technical article on "running agentic skills with Gemma 4 at the edge" — LiteRT-LM is the infrastructure layer for that direction. If you're building mobile AI or edge deployments, this is one of the most important Google open-source projects to track right now.

📈 #11 — nvim-treesitter/nvim-treesitter｜13K-Star Plugin Archived, Neovim Community Shaken

"Nvim Treesitter configurations and abstraction layer"

+154 ★ this week｜13,647 total｜Tree-sitter Query｜Apache-2.0

The lowest weekly growth in this list, yet the hottest open-source story on HN. nvim-treesitter was archived by its maintainer on April 3, 2026, sparking a 176-point HN discussion.

Why it was archived: The maintainer completed a full rewrite targeting Neovim 0.12 in March 2026 and clearly documented "0.11 users: use the frozen master branch." Despite this, a large number of 0.11 users continued opening issues and PRs demanding backward compatibility. The maintainer ultimately archived the repo to end the drain.

This triggered a broader discussion: where is the line for open-source maintainers? Archiving a plugin with 12K forks and 60+ language grammars — is that a maintainer's victory or a community failure? Top-rated HN comments mostly sided with the maintainer. In practice, the community has already forked to neovim-treesitter/nvim-treesitter and tree-sitter-manager.nvim.

Neovim users: update your plugin configs and stop depending on the archived repo.

Weekly Spotlight — Top New Repos Top 15

🆕 #1 — MemPalace/mempalace｜"Highest-Scoring AI Memory System Ever" Faces Benchmark Scrutiny

"The highest-scoring AI memory system ever benchmarked. And it's free."

43,367 total stars｜Python｜MIT｜2026-04-05

MemPalace launched on April 5 and crossed 20K stars within 48 hours — the undisputed leader among newly born repos this week. Part of the story: one of its co-founders is Hollywood actress Milla Jovovich (Resident Evil franchise), and that "celebrity startup" angle dramatically amplified media attention.

Technically, MemPalace uses a "full verbatim storage + vector search" architecture: every conversation is stored word-for-word (no AI summarization), with local vector retrieval via ChromaDB + SQLite, zero API costs, fully offline. It claims 96.6% raw and 100% hybrid scores on the LongMemEval benchmark.

However, HN discussion (66 pts) and independent evaluators challenged that "100%": the perfect hybrid score was achieved by targeted fixes for specific failure cases, not general performance improvement. Methodological choices like top_k were also questioned as favorable to their own test setup. HackerNoon's headline was blunt: "devs shredded its benchmarks."

What this means for you: MemPalace's "full verbatim storage + local vector search" architecture is a genuinely reasonable design choice, especially for privacy-sensitive scenarios. But form your own judgment on the benchmarks — run it against your own test cases before adopting.

🆕 #2 — alchaincyf/nuwa-skill｜Distill Anyone's Thinking Into an Executable Skill

"Distill how anyone thinks."

8,453 total stars｜Python｜MIT｜2026-04-05

nuwa (女媧, the Chinese mythological creator goddess) is a Claude Code Skill whose core capability is extracting a "cognitive operating system" from a public figure's publicly available material — not copying their catchphrases, but extracting mental models, decision heuristics, and expressive DNA so that AI can answer new questions using that person's thinking framework.

In practice, nuwa deploys 6 parallel agents to simultaneously research a target person from different angles (written works, podcasts, social media, critics' perspectives, decision records, life timeline). A claim must pass three tests — cross-domain consistency, predictive power, and exclusivity — before it gets recorded as a mental model.

Existing distillation examples include Steve Jobs, Paul Graham, Zhang Yiming, Karpathy, Ilya Sutskever, Charlie Munger, Naval Ravikant, and Nassim Taleb.

This repo ignited this week's persona distillation wave — the same week saw zhangxuefeng-skill, khazix-skills, tong-jincheng-skill, and the awesome-persona-distill-skills aggregator all emerge.

🆕 #3 — garrytan/gbrain｜A Personal Knowledge Management Brain for Hermes Agent

"Garry's Opinionated OpenClaw/Hermes Agent Brain"

6,210 total stars｜TypeScript｜MIT｜2026-04-05

gbrain is a highly personalized Hermes Agent configuration repo — think of it as "an engineer open-sourcing their personal thinking toolkit." It integrates a knowledge graph, personal memory system, and decision support tools, designed as a memex for knowledge workers.

🆕 #4 — alchaincyf/zhangxuefeng-skill｜Zhang Xuefeng's College Admissions Decision Framework as a Claude Code Skill

"张雪峰.skill — 高考志愿/考研/职业规划的实战思维框架。由女娲.skill生成。"

5,269 total stars｜MIT｜2026-04-05

Zhang Xuefeng is a well-known Chinese college admissions consultant known for his pragmatic, ground-level analysis. This skill uses the nuwa-skill framework to distill his decision logic around college major selection and career planning into an executable Claude Code Skill.

Its viral traction reveals an important use case for Persona Distillation: making scarce domain expertise (college advising, legal guidance, financial advice) accessible to more people at lower cost. Questions of accuracy and compliance remain open, but the demand is real.

🆕 #5 — farzaa/clicky｜An AI Tutor That Watches Your Screen as You Work

3,936 total stars｜Swift｜MIT｜2026-04-07

Clicky is a macOS menu bar app that describes itself as "the person who stands next to your screen, watches your pixels, and points to the answer when you hesitate." Press push-to-talk: it takes a screenshot and records audio simultaneously, sends both to AssemblyAI for transcription and Claude for analysis, reads the response aloud via ElevenLabs TTS, and points to the relevant screen location with a transparent cursor overlay.

An interesting architectural choice: a Cloudflare Worker proxy keeps API keys server-side so the desktop client never holds sensitive credentials. The community has already built a Windows version (Electron + TypeScript).

🆕 #6 — xixu-me/awesome-persona-distill-skills｜A Curated List of Persona Skills

"同事.skill, 女娲.skill, 前任.skill… Curated list of Agent Skills centered on people, relationships"

3,404 total stars｜JavaScript｜CC0-1.0｜2026-04-06

This awesome list aggregates Agent Skills centered on people and relationships — colleague.skill, nuwa.skill, ex.skill, and more. The fact that a curated list itself attracted 3.4K stars signals that this sub-ecosystem has grown large enough to justify aggregation tools.

🆕 #7 — LaurieWired/tailslayer｜Eliminating a RAM Latency Problem That's Been Around Since the 1960s

"Library for reducing tail latency in RAM reads"

2,091 total stars｜C++｜Apache-2.0｜2026-04-05

tailslayer is the only newly born repo this week outside the AI/Skills ecosystem that still earned 110 points on HN — pure systems engineering.

It targets DRAM refresh stalls: modern DDR DRAM must periodically pause all reads and writes to perform refresh cycles, a mechanism that has existed since IBM's original 1960s DRAM design. This causes tail latency (p99.99) to be dramatically higher than average latency.

tailslayer's approach is "hedged reads": replicate data across multiple independent DRAM channels (each with uncorrelated refresh schedules), then issue simultaneous requests to all channels on read and use whichever responds first. On AMD EPYC Turin (12 memory channels), tail latency drops by up to 89%. The cost: you need to replicate your entire working set, doubling memory usage.

Tom's Hardware's headline called it progress "with severe downsides," but HN consensus is that it has real value for high-frequency trading and real-time systems. LaurieWired has a full technical explainer on Twitter/X.

🆕 #8 — alchaincyf/hermes-agent-orange-book｜A Complete Guide to Hermes Agent in Chinese

"Hermes Agent 从入门到精通 · 橙皮书系列 · Nous Research 开源 AI Agent 框架实战指南"

2,088 total stars｜2026-04-08

This is a Chinese-language hands-on guide to Hermes Agent ("Orange Book" series), written by alchaincyf — the same author as nuwa-skill. Its existence is itself a signal: when community-generated documentation starts forming organically around an open-source framework, its ecosystem has crossed a maturity threshold.

🆕 #9-10 — KKKKhazix/khazix-skills and hotcoffeeshake/tong-jincheng-skill｜The Long Tail of Persona Distillation

khazix-skills (★1,709): A collection of AI Skills packaged around the "digital life of Kha'Zix" (a League of Legends character), a game-character framing for a practical toolkit
tong-jincheng-skill (★1,590): A Skill analyzing interpersonal relationships from the perspective of Tong Jincheng, a well-known Chinese emotional wellness content creator

These two repos represent the long tail of the persona distillation wave — it's not just prominent thinkers. Any public figure with a distinctive "thinking style" can become a distillation target.

🆕 #11 — yizhiyanhua-ai/fireworks-tech-graph｜Production-Quality Technical Diagrams via Claude Code

"Claude Code skill for generating production-quality SVG+PNG technical diagrams."

1,530 total stars｜Python｜MIT｜2026-04-10

Supports 8 diagram types (architecture, sequence, flowchart, etc.) and 5 visual styles, with deep domain knowledge in AI and agent systems. If you're writing technical documentation or RFCs with Claude Code, this skill is worth trying.

🆕 #12 — fikrikarim/parlor｜Fully Offline Real-Time Multimodal Voice AI

"On-device, real-time multimodal AI. Have natural voice and vision conversations with an AI that runs entirely on your machine."

1,417 total stars｜HTML｜Apache-2.0｜2026-04-05

Parlor uses Gemma 4 E2B + Kokoro TTS to deliver fully local real-time voice + vision conversation, running entirely on your machine. Supports Apple Silicon, depends on LiteRT-LM for inference. It's a community-side demonstration of what the google-ai-edge ecosystem makes possible.

🆕 #13 — mattmireles/gemma-tuner-multimodal｜Fine-Tune Gemma 4 Multimodal on Apple Silicon

1,229 total stars｜Python｜MIT｜2026-04-07

Supports multimodal fine-tuning of Gemma 4 (including 3n) on MacBook/Mac Studio using PyTorch + Metal Performance Shaders, covering audio, image, and text modalities. Lowers the barrier for local LLM fine-tuning significantly.

🆕 #14 — nashsu/llm_wiki｜Let LLMs Automatically Build Your Personal Knowledge Wiki

"Cross-platform desktop app that turns your documents into an organized, interlinked knowledge base."

907 total stars｜TypeScript｜2026-04-08

llm_wiki takes a different philosophy from traditional RAG: rather than retrieving and answering from scratch each time, it has the LLM continuously build and maintain a persistent wiki from your documents incrementally. This "living knowledge base" model is closer to how people actually organize knowledge over time.

🆕 #15 — phuryn/claude-usage｜A Local Dashboard for Tracking Claude Code Token Usage

"A local dashboard for tracking your Claude Code token usage, costs, and session history."

878 total stars｜Python｜MIT｜2026-04-07

Addresses a real and widespread pain point: Claude Code Pro/Max subscribers only see a progress bar, with no granular token or cost breakdown. claude-usage provides a local dashboard with full visibility into per-session token consumption and cost estimates.

Three repos appeared on both the weekly and monthly trending charts this week (🔁):

NousResearch/hermes-agent: Growing continuously since July 2025, this week's single-week +32K is a new all-time high. Monthly persistence means it's not a one-off spike — it reflects genuine sustained adoption.
google-ai-edge/gallery: Has held monthly trending for consecutive weeks, serving as Google's primary showcase window for its edge AI strategy.
google-ai-edge/LiteRT-LM: As the underlying execution engine for both gallery and parlor, its monthly presence signals that developers are actually deploying this on-device AI stack in real projects.

All three 🔁 repos are infrastructure-layer projects, not one-shot viral hits. That's an important trend signal.

Weekly Trend Insights

Persona distillation became the most visible new ecosystem explosion this week

Last week nuwa-skill appeared; this week it directly triggered a wave of derivatives: zhangxuefeng-skill, khazix-skills, tong-jincheng-skill, and awesome-persona-distill-skills all emerged in the same seven-day window, alongside andrej-karpathy-skills (already at 16.5K stars) experiencing another surge. This pattern makes clear that "distilling someone's thinking style into an executable Skill" has evolved from individual hacks into a distinct repo category. Worth noting: at its core, this wave is about "personalized knowledge distribution" — it changes how knowledge is packaged and transmitted, but also opens real questions around accuracy and intellectual property.

Hermes Agent is becoming the host platform for the Skills ecosystem

Hermes Agent took the weekly growth crown, but more importantly it's taking on a new role: all Persona Skills, Karpathy Skills, nuwa-skills, and similar projects need an agent framework to run on. Hermes Agent's openness (200+ LLMs, multi-platform deployment) makes it the natural choice. The appearance of hermes-agent-orange-book signals that community-generated ecosystem documentation is forming organically.

On-device AI infrastructure is quietly taking shape

Google (gallery + LiteRT-LM), NVIDIA (PersonaPlex), and the community (parlor, gemma-tuner-multimodal) all advanced this week simultaneously. This isn't coincidence — it's the industry collectively validating the proposition that "AI doesn't require the cloud." parlor delivers fully offline real-time voice + vision conversation using Gemma 4 + LiteRT-LM on a MacBook. That's working reality today, not a demo. For developers: if your use case involves privacy requirements or low-latency constraints, now is the right time to start evaluating on-device AI stacks.