by not4humans
MuscleMemory: where trial-and-error compiles into skills you can trust.
- LLMs gave us raw intelligence.
- RAG gave them memory.
- Agents gave them tools.
- MuscleMemory gives them skills.
Today, agents re-solve the same problems from scratch every time. Each run is clumsy, fragile, and costly. With MuscleMemory, they can practice, make mistakes, and compile experience into deterministic skills that can be applied again and again.
Reliable. Auditable. Local-first. One line of code.
from not4humans.musclememory import MuscleMemory
agent = MuscleMemory.wrap(my_agent, tools=[term_exec, fs_patch], llm_client=openai_chat)
result = agent.run("Deploy latest to staging")
# trial and error → skill → reliable outcome- Deterministic execution — every skill runs with explicit preconditions, budgets, and replay.
- Reliability — once learned, skills execute predictably.
- Drop-in adoption — wrap your agent in one line, no rewrites.
- Local-first — logging, CI, and skill mining all work offline.
- Governance (optional) — promote skills across teams, enforce policy, and audit at scale.
- 📖 Quickstart Guide
- 📚 Skill Manifest Schema
- 🛠️ Integration Checklist
- SDK core (wrap, session, deterministic gate, plan runner)
- Local registry + miner + compiler
- LangGraph / LangChain / Semantic Kernel adapters
- Control-plane client (policy, promotions, telemetry)
- Example skills library
👉 From prompts → memory → tools → skills.
MuscleMemory is the obvious next steip.