跪拜 Guibai
← All articles
Agent · TypeScript · News

Agent Skills and Memory Form a New Infrastructure Layer as GitHub Trending Diverges from Growth

By 苏灿烤鱼 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent infrastructure is splitting into three concrete layers—long-term memory, file workspaces, and task-loop state—and the projects building them are gaining traction fast. Relying on GitHub's trending rank alone now hides which projects are actually accelerating.

Summary

Two Agent skill repositories—addyosmani/agent-skills and mattpocock/skills—entered the top four on the same day, encoding senior engineering workflows into executable Agent processes rather than broad prompts. Meanwhile, TencentDB-Agent-Memory returned to #1 after six days on the list, and Cloudflare's computer runtime posted the day's highest growth at +2,690 stars despite ranking only third. Together, memory, workspace, and loop-state projects are assembling the three-layer foundation needed to run Agents continuously.

The ranking itself is becoming a less reliable signal. pdf-inspector gained +1,194 stars but landed at #13; DeepSeek-Reasonix added +894 at #11. Both outgrew the #1 spot. GitHub's algorithm produces a snapshot that obscures momentum, making it necessary to track rank, daily growth, and historical trajectory side by side.

Classic projects like Google's Guava and AutoGPT returned to the list with negligible daily growth, scattering the day's structure. The through-line remains Agent tooling, but the returning entries resist being folded into a single narrative.

Takeaways
— TencentDB-Agent-Memory returned to #1 after a six-day trajectory of #13→#10→#5→#1→#3→#1, gaining +1,053 stars.
— Cloudflare's computer runtime posted the day's highest growth at +2,690 stars but ranked only #3.
— Two Agent skill repositories—addyosmani/agent-skills and mattpocock/skills—entered the top four simultaneously, packaging engineering workflows as executable Agent processes.
— addyosmani/agent-skills ships 24 skills and 8 commands covering the full development cycle and recently fixed four Agent personas that silently failed to load in Claude Code.
— mattpocock/skills debuted with +2,002 stars by open-sourcing a personal `.agents` directory as an engineering skill collection.
— pdf-inspector gained +1,194 stars but ranked #13, and DeepSeek-Reasonix gained +894 at #11—both outgrew the #1 project.
— Google's Guava entered at #7 with only +35 stars, and AutoGPT at #9 with +28, scattering the day's thematic coherence.
— tirth8205/code-review-graph persists a local codebase relationship graph so review tools read only relevant context, reducing what an Agent must parse.
Conclusions

Agent skills are productizing senior engineering judgement—requirement clarification, planning, testing, review, and release—into executable workflows, not just better prompts. This shifts the bottleneck from model capability to process fidelity.

The three-layer Agent infrastructure emerging from these projects (memory, workspace, loop state) mirrors the stack that made web applications reliable: persistence, file systems, and job queues. The pattern is repeating for autonomous agents.

GitHub's trending algorithm is producing rankings that invert growth signals. A project with +2,690 stars ranks below one with +1,053. Anyone using the trending page as a discovery feed is seeing a distorted picture of what is gaining momentum.

Classic repositories like Guava and AutoGPT returning with negligible growth suggests the algorithm has a recency or baseline-star bias that pulls in established projects regardless of current activity, diluting the list's signal for new infrastructure work.

Concepts & terms
Agent Skill
A packaged, executable workflow that encodes a specific engineering task—such as requirement clarification, testing, or release—so an AI agent can follow it reliably, rather than relying on a broad natural-language prompt.
Agent Memory
A persistence layer that stores an agent's long-term context, conversation history, and learned preferences across sessions, enabling continuous operation without losing state between runs.
Agent Workspace
A file-system environment managed for an agent, providing a sandboxed space where it can read, write, and organize files during a task, analogous to a developer's project directory.
Task Loop State
The runtime mechanism that tracks an agent's progress through a multi-step goal, allowing it to pause, resume, and iterate without repeating completed work or losing its place.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗