Symbio

A personal AI agent that runs 100% locally and fine-tunes itself on what it learns.

MLX + LoRA on Apple Silicon self-correction learns skills on the fly remembers web research browses on its own RAG memory sandboxed tools cron reminders

The model itself trains locally on your Mac (MLX is Apple-Silicon-only), so this page demos the agent's actual machinery: every panel below runs the real Python modules from the GitHub repo, in your browser, via Pyodide.

🛡 Golden-set regression checks

Before and after every LoRA training pass, Symbio runs a fixed battery of prompts that exercise the contracts seeded into training data: identity, the right tag for each task, and recovering from common tool mistakes. This panel runs the real symbio/app/golden.py checks on mocked correct replies.

…

🏷 The tag language

Symbio's model acts through tags in its replies. Paste a reply and run the real parser (symbio/app/tooling.py) to see the extracted tool calls and what the user actually sees.

Model reply

…

🔁 Self-correction when hallucinating

Correct Symbio and it captures the exchange as a mistake note; at the configured threshold it digests them into boosted training samples and runs a short LoRA pass — the wrong answer is trained away. This panel runs the real detector and miner (symbio/app/learn.py).

Original question

Assistant's wrong answer

Your correction

Assistant's corrected answer

…

🌐 Remembering web research

When a turn used web tools, the final answer is auto-saved as a Learned: note — unless it's ephemeral (weather, news, prices), trivial, or already known. Real filter, real note.

Question

Answer the agent found

…

🧩 Skills on the fly + RAG recall

A working multi-step approach gets saved with <skill>; the keyword retriever (rag.py) pulls it back into context when a similar task appears — and digest bakes it into the weights.

Skill name

Steps

…

Now ask something related — watch retrieval bring the skill back

…

🧬 How the growth loop fits together

 you chat ──► agent loop (tags → sandboxed tools → observations)
                 │
                 ├─ <note> / <skill> / auto research notes ──► notes/  ──► RAG (instant recall)
                 ├─ corrections ──► notes/mistakes/ ──► boosted samples
                 └─ <memory> / <profile> ──► always-in-context stores
                                   │
                       <digest /> + <train />
                                   │
                     MLX LoRA fine-tune (local, Apple Silicon)
                                   │
                   the model itself gets permanently better

See the actual system prompt that teaches the model its tags

loading…

Run the real thing

Everything above, plus the local LoRA training loop, live browsing, cron scheduling, and permission-gated sandbox — on your own Mac, fully offline-capable.

git clone https://github.com/huyedits/Symbio && cd Symbio
./setup.sh
python main.py

⭐ github.com/huyedits/Symbio

Built with the actual Symbio source — the Python running in this page is the same code that runs the agent.
MIT-style licensed · Contribute on GitHub