GitHub's monthly trending page fails the moment you try to filter for 'AI.' A language filter defines the set for Rust or Go, but open source AI spans five major languages and dozens of job functions. A new snapshot from October 16, 2026, ignores the language filter entirely, instead grouping repositories by the job they perform. The result is a clearer picture of where builders are actually spending their time: not on model weights, but on the plumbing that connects agents to memory, workflows, and local runtimes.

The Personal Agent Race Is Won by OpenClaw

OpenClaw leads the personal assistant category with 392,126 stars, far outpacing Nous Research's Hermes Agent at 254,122. This isn't just about raw popularity; it reflects a shift toward local-first automation. OpenClaw is a TypeScript-based runtime that executes tasks on your own OS, while Hermes Agent focuses on accumulating skills and context over time. AutoGPT remains relevant with 187,598 stars, but its license has split: the core is MIT, while the platform folder is under Polyform Shield, restricting commercial value. Firecrawl, with 191,908 stars, is technically a web-data tool, but its inclusion highlights that agents are starving for clean, structured web inputs.

Coding Assistants Are Fragmenting by Workflow

The coding agent space is dense, with OpenCode taking the top spot among fully open tools at 213,922 stars. It is followed closely by vendor CLIs that happen to be open source: OpenAI's Codex (129,395 stars, Rust) and Google's Gemini CLI (107,389 stars, TypeScript). The key differentiator here is not the model, but the integration. OpenHands (91,004 stars) offers a full task loop with sandboxes, while Cline (70,370 stars) and Goose (55,324 stars) serve as SDKs and workstation agents respectively. Builders must choose between a self-hostable platform, a vendor-linked CLI, or a swappable local agent. Running multiple coding agents per repository creates configuration debt that rarely pays off.

Memory Layers Are the New Bottleneck

As agents become more autonomous, the ability to remember is becoming the primary infrastructure concern. claude-mem leads this group with 101,052 stars, integrating with almost every major coding agent to compress session history. Mem0 (67,243 stars) and Letta (25,156 stars) represent the service-oriented approach, offering memory APIs that sit outside the coding tool. Beads (27,917 stars) takes a different tack, storing tasks and plans as files that travel with the git branch. The split is clear: session-based memory for coding tools versus application-based memory for broader agents. Choosing the wrong layer means your agent either forgets its last commit or fails to remember your user's preferences.

Key Takeaways

  • OpenClaw and Superpowers are the most starred AI projects, indicating a massive interest in agentic frameworks over standalone models.
  • License complexity is rising: AutoGPT and n8n have split or restrictive licenses that impact commercial deployment.
  • Memory is no longer an afterthought; claude-mem and Mem0 are critical infrastructure for stateful agents.
  • Vendor CLIs like Codex and Gemini CLI are gaining traction as open-source clients for hosted models, bridging the gap between local tooling and cloud intelligence.

The Bottom Line

Stop chasing the highest star count for a model; chase the highest star count for the workflow that solves your specific bottleneck. The real innovation in October 2026 is in the glue: memory, harnesses, and gateways. The shift from 'which model is best' to 'how do I orchestrate this' is complete. Infrastructure is the new differentiator.