Today's GitHub trending board tells a coherent story: the AI agent ecosystem is maturing past "demo-worthy chatbot" and into something that looks more like real software infrastructure. The projects gaining traction are the ones tackling memory, cross-platform deployment, cost control, and resilient data collection — the unglamorous plumbing that makes agents actually usable in production.
NousResearch/hermes-agent: An Agent That Actually Learns
Nous Research's hermes-agent is an open-source framework for building AI agents that improve autonomously over time. Unlike most agent frameworks where sessions are stateless, Hermes includes a built-in learning loop — the agent creates and refines its own skills from experience, maintains memory across sessions, and can run continuously on infrastructure as modest as a $5 VPS. It works across Telegram, Discord, Slack, WhatsApp, Signal, and CLI from a single codebase, and supports over 200 models via OpenAI, Anthropic, and OpenRouter. At 1,735 stars today, it's the standout new entry on the board.
github.com/NousResearch/hermes-agent
supermemoryai/supermemory: Persistent Memory as an API
supermemory takes a focused approach to one specific gap in AI applications: most assistants forget everything the moment a conversation ends. The project extracts facts automatically, builds and maintains user profiles, and delivers relevant context at query time through a hybrid retrieval system that blends standard RAG with personalized memory. It connects to sources like Google Drive, GitHub, and Notion, and integrates with Vercel AI SDK, LangChain, and Claude. With 600 stars today, it's positioning itself as the standalone memory layer developers would otherwise have to build themselves.
github.com/supermemoryai/supermemory
JuliusBrussee/caveman: 65% Fewer Tokens, Same Results
caveman is a Claude Code skill that takes a blunt approach to token costs: just tell the model to stop using so many words. The skill instructs the agent to drop filler language, use sentence fragments, and communicate with terse precision — the project's own description puts it as "why use many token when few token do trick." It claims 65–75% token reduction, offers multiple compression levels (lite, full, ultra, and a wenyan classical Chinese variant), and supports 30+ agents including Codex, Gemini, and Cursor. The 471 stars it earned today suggest developers find it genuinely useful despite the jokey framing.
github.com/JuliusBrussee/caveman
D4Vinci/Scrapling: Web Scraping That Adapts
Scrapling is a Python scraping framework designed for the messiness of real websites — layout changes, anti-bot protections, and the need to scale from a single request to a full crawl. Its standout feature is adaptive element tracking: if a site restructures its pages, Scrapling can relocate the elements you were targeting rather than simply breaking. It also includes stealth and full browser automation modes for sites behind Cloudflare and similar protections, and ships an MCP server for AI tool integration. With 1,067 stars today, it's the most practical data-gathering tool on the board for developers building AI pipelines that depend on live web content.