What if your AI didn’t just respond to you…
What if it grew?
What if it could forge new abilities, install them, swap them, and persist them — all while running?
That’s exactly what the LivinGrimoire SkillForge does.
This isn’t a plugin system.
This isn’t a tool‑calling wrapper.
This is LLM digivolution — a self‑extending AI organism that upgrades itself on demand.
And yes, it runs in Python.
🔗 The Engine Behind It All
All of this is powered by the LivinGrimoire software design pattern, a modular AI runtime that treats abilities as hot‑swappable “skills.”
👉 GitHub: https://github.com/yotamarker/LivinGrimoire
If you’ve never seen an AI runtime that behaves like an operating system for skills… buckle up.
⚡ What “LLM Self‑Digivolution” Actually Means
It means:
- The AI listens for explicit capability requests
- It calls the LLM only when needed
- The LLM writes a new Python Skill class
- The engine hot‑loads that class at runtime
- The AI gains a new ability instantly
- If the skill slots are full, it evicts the least useful one
- Protected skills (memory, identity, wellbeing) can’t be removed
- Everything persists across sessions
This is not theoretical.
This is not a prompt trick.
This is runtime evolution.
🧩 SkillForge Algorithm (Bullet‑Pointed for Devs)
Trigger Phase
- Listen for explicit phrases like:
- “can you X”
- “i wish you could X”
- “why can’t you X”
- Match using local regex (zero token burn)
Forge Phase
- Build a natural‑language description of the requested ability
- Send it to the LLM with a strict skill‑generation system prompt
- Extract the Python code block
- Extract the class name
- Reject duplicates using persistent history
Hot‑Load Phase
-
exec()the generated code in a sandboxed namespace - Instantiate the new Skill
- Wire it into the shared Kokoro (brain)
Equip Phase
- Read the skill’s notes to detect protected categories
- If capacity not full → add skill
- If full → ask LLM which skill to evict (except protected ones)
- Swap skills using APSkillSwapper
- Persist:
- history
- equipped list
- protection flags
Runtime Phase
- Skill becomes active immediately
- No restart required
- No rebuild required
- No manual coding required
This is live evolution.
🛠️ SkillForge Features (Bullet‑Pointed)
- 🔍 Local trigger detection (no LLM call until needed)
- 🧠 LLM‑generated Python Skill classes
- ⚡ Hot‑loading via exec()
- 🔄 Automatic skill installation & activation
- 🧹 LLM‑guided eviction when full
- 🛡️ Self‑preservation protection keywords
- 🧬 Persistent memory of forged skills
- 🧩 APSkillAdder + APSkillSwapper integration
- 🧵 Thread‑safe forging pipeline
- 🧱 Lobe‑based architecture (logic, TTS, STT, sensors, vision)
- 🧰 SkillNotes‑driven introspection
- 🧯 Fallback eviction if LLM reply malformed
- 🧿 Console logging of generated code
- 🧠 History‑aware forging (no duplicates)
This is the closest thing to a self‑modifying AI runtime you can build today without going full sci‑fi.
LLM Self‑Digivolution: The Plug‑and‑Play Skill That Lets AI Evolve New Abilities in Real Time
Code:import re import threading import requests from LivinGrimoirePacket.LivinGrimoire import Skill, Brain from AlgParts import APSkillAdder, APSkillSwapper #
yotamarker.com
Top comments (0)