feat: add kinthai-self-improving-user (per-user learnings isolation for multi-user agents)

[kinthaiofficial]

Skill: kinthai-self-improving-user

Repo

https://github.com/kinthaiofficial/kinthai-self-improving-user

Category suggestion

🤖 AI & LLM Tools (or new category: Agent Memory)

Description

User-level self-improvement skill for OpenClaw agents serving multiple users. Builds on @pskoett's Self-Improving Agent but isolates learnings per user_id instead of per workspace.

.learnings/
  ├── _global/     ← shared across all users (universal patterns)
  ├── 10000002/    ← User A's corrections, preferences, profile
  ├── 10000003/    ← User B's corrections, preferences, profile
  └── _meta/       ← promotion logs, stats

Three behaviors that aren't in pskoett's original:

1. Per-user isolation — user_id is extracted from OpenClaw session keys at bootstrap, learnings are written under <user_id>/
2. Two-layer reading — agent reads both _global/ and <user_id>/ and concatenates them into the system prompt
3. Pattern promotion — when 3+ different users independently trigger the same learning, it auto-promotes to _global/

When you'd want this

• Single OpenClaw agent serving multiple users (marketplace, family bot, team assistant)
• You want the agent to learn user-specific preferences without leaking them across users
• You also want commonly-correct patterns to become universal automatically

License

MIT

Why I think it fits the curated list

The existing master-skills skill is about secrets management; this is the equivalent for learnings/profiles. Together they cover the two storage layers an OpenClaw agent needs to be production-multi-user-ready.

(Disclosure: this skill is maintained by KinthAI — we run it in production on our platform. The skill itself is fully open-source MIT and works in any OpenClaw deployment.)

Happy to open a PR with the directory + entry in the right section if you'd prefer that path.