A Harness engineering practice, I open sourced it!
It’s quite interesting.
After using Claude Code for two months, I discovered a crazy problem: it keeps making the same mistakes every time.
Tell it "Feishu API uses index=0", and next time use index=1.
Tell it "git push requires a proxy", and the direct connection times out next time.
AI has no memory.
Every session is an intern’s first day on the job.
This isn't Claude's problem.
This is an architectural issue.
LLM itself has no persistent memory.
CLAUDE.md can solve part of the problem, but you have to write it yourself, maintain it yourself, and put everything in one file.
What is really needed is: a memory system that is classified by skill, automatically accumulated, and automatically injected.
I made an open source project to solve this problem: fireworks-skill-memory.
The core idea is very simple: when using a skill, the "historical lessons" of this skill are automatically injected into Claude.
After the session is over, the new lessons are automatically extracted and saved.
The next time you use it, it will be automatically better.
It is hung on the Harness layer of Claude Code.
Many people don't know what Harness is - it is the orchestration layer between the model and the real world in Claude Code, responsible for tool invocation, Hook life cycle, and Skill loading.
Fireworks-skill-memory only uses two Hook points: PostToolUse: Detect Skill being read → Inject the corresponding KNOWLEDGE.md Stop (async): After the session ends → Haiku refines new lessons → Updates the file.
It has zero impact on your workflow.
Totally in the background.
GitHub: MIT license, PRs welcome.
If you are also using Claude Code, give it a try.
The installation method is ridiculously simple: Say in Claude Code: "Help me install fireworks-skill-memory from here: After the first session, go to ~/.claude/skills/ to see what it automatically writes - each skill has refined common pitfalls and best practices, ready to use out of the box.
Privacy Note: All knowledge files are stored locally under ~/.claude/skills/.
No content is uploaded to the cloud.
Refining is using claude-haiku (your own API key).
The content of your conversations never leaves your machine.
The essence of this project is to outsource the experience of using AI from the human brain to the machine itself.
You no longer need to remember "which API has pitfalls".
You no longer need to manually take notes in CLAUDE.md.
AI will accumulate experience for itself, and the more it is used, the smarter it will become.
This is how AI native works.