A free coding agent that writes its own notes — so it gets better the more you use it.
Prime Agent is open source (MIT, ~22.8k stars). You bring a free model key; it does the work.
1 · Run one line
Ori → pick a free model → one command. No card.
2 · Give it a small real task
“build a tiny script” · “fix this error” · “explain this file”.
Its only tool is a small Python workspace — it writes code to do the job.
3 · It hires helpers
It opens copies of itself for parts of the work. Insiders: rlm.spawn() hands back a job ticket first, the answer later.
4 · It saves the lesson
After the run it writes notes and skills into its own memory — a reversible edit. That is the “self-improving” part.
task
│
▼
code ──► helpers
│
▼
memory
│
↺──► next run, better
5 · The cheeky bit
A four-file MIT cousin, telos, goes further.
It rewrites its own source and restarts mid-chat — it even writes its own first system prompt.
Keep it in a throwaway box.
Start free: one command, one free key, one small real task.
What it is, every free door, the bolt-ons — and the honest bits
What it is — insider words with plain meanings
• Prime Agent — ~22.8k stars · MIT — use it, change it, share it · current release
• the maker’s plain explanation · the paper — how the memory loop works
• insiders call the setup RLM. It copies itself into helpers (rlm.spawn) and saves what worked into notes and skills that outlive the chat — the continual harness. Each edit is logged and reversible.
Free doors — pick one
• Ori — logs you in, no key to paste · one-line install
• OpenRouter free models — 50 calls a day at $0 · limits · live provider table · llm-keypool — free keys on rotation
• Google AI Studio — big daily allowance · Groq free — very fast, short chats · Ollama — no account, no key, runs on your own machine · Codespaces — 120 free hours in a browser
• direct install + provider guide — the hands-on route
Bolt-ons
• browser-use 1.0 — the same one-helper idea for web pages · SkillLib — 20,000+ skills · qai.io skills — 14,000+ more · agentic-awesome-skills — 3,000+ more
• rat and pythond — share one live Python space across programs · marimo pair — helpers linked by visible variable lines
• Langfuse — see every helper call · E2B — a throwaway machine for risky code · more sandbox options
• mengram — memory that learns from mistakes · primecord-agent — always-on chat version · A2A + orxhestra — helpers from other tools · ACE — the playbook method behind the memory
• community setup guide · memory atlas — first-run fixes and field notes
Honest bits, with receipts
• the 95.5% score you may have seen is a self-reported efficiency figure on an easier public set, not an official result — plain explainer · its own run-by-run table · why the model matters more
• /refine edits notes and skills only · free keys suit short chats; long runs cost real money · the agent runs code with your permissions — use a throwaway folder.
Send one message. Watch it remember. ![]()

!