Show HN: Symbio self fine-tuning AI loop

Symbio, a local AI assistant that learns from user corrections and fine-tunes itself without cloud or subscriptions, was released on GitHub by developer HuyEdits. The system runs on Apple Silicon using MLX and Metal performance shaders, requires 16GB unified RAM, and features a Mixture of Agents mode where a headmaster delegates tasks to smaller worker models, with fine-tuning triggered after 5+ repeated mistakes. Symbio supports skills, adapters, Telegram, and local file operations, aiming to reduce repetitive instructions.

Local Ai that learns from your corrections. No Cloud, No subscriptions. Symbio takes notes and can construct it into training data to fine-tune itself - so you can stop repeating yourself. | https://huggingface.co/spaces/HuyEdits/symbio-demo GitHub https://github.com/huyedits/Symbio | Try it now quick-start Symbio develops as you tell it what to do in repeat. Try the interactive demo — the agent's real tag parser, self-correction miner, research memory, and RAG retriever running in your browser: https://huggingface.co/spaces/HuyEdits/symbio-demo https://huggingface.co/spaces/HuyEdits/symbio-demo Symbio has a MOA Mixture of Agents mode. Instead of fine-tuning one big model for every task, the headmaster delegates bounded sub-tasks to smaller worker models via tool calls. The worker executes, and if it fails it returns to the headmaster for guidance. Once it works, a note is saved for both sides. If the same mistake repeats past the configured threshold, both the worker and the headmaster are fine-tuned: the worker learns how to execute the task, and the headmaster learns how to delegate it more efficiently. Symbio can learn skills on the fly. A skill starts as a simple markdown note with step-by-step instructions. As errors and corrections accumulate, they are logged in a hidden .md.health.jsonl sidecar so the note itself stays clean and readable. Once the mistake threshold is reached, the collected examples are fed into a LoRA fine-tune that creates a dedicated worker adapter for that skill — one adapter = one skill. Adapters are hot-swappable and can be archived if unused. Use /new-skill <name or symb skill new <name to create one, /skill-adapters to list them, and /archive / /restore to manage idle notes and adapters. - Chat through a local CLI or a Telegram bot. - Save facts and notes as markdown files in notes/ . - Read, write, search, and patch files inside the project directory. - Run sandboxed shell commands and short Python snippets. - Check email via IMAP/SMTP when configured . - Digest notes into training data and fine-tune a LoRA adapter on the fly. - Persist every conversation turn to JSONL and an SQLite store. Please star this project too. It would help me out SO SO much. :3 <3 Symbio for now runs on Apple Silicon using MLX and Metal performance shaders Recommended Unified RAM requirements 16gb the program itself takes 8 but overhead and expansion so comfortably would be 16 Minimum architecture any m-series chip ideally. if you can let me know if it works for m1,m2,m3,m5 and the pro and max variants Create a virtual environment and install dependencies python -m venv venv source venv/bin/activate pip install -r requirements.txt Install symbio / symb as system-wide commands. Editable install links the current repo so code edits take effect immediately: pip install -e . First launch runs an interactive setup wizard: - set names, pick a model preset, choose speed mode - toggle browser, web search, MOA dispatch, Telegram - add Telegram bot token and allowed chat IDs symbio Or the short alias: symb Re-run the wizard anytime with: symb setup You talk to the AI — Ask it anything It makes mistakes — Sometimes gets it wrong You correct it — "No, it's actually..." It learns — Saves the correction After 5+ corrections, it fine-tunes itself Next time: it gets it right : That's it. No manual training. No API calls. All local. If you prefer an isolated, non-editable install e.g. with pipx : pipx install . or, from any directory containing this repo: pipx install /path/to/agi Make sure ~/.local/bin or your pip/pipx bin directory is on your PATH . On first run, Symbio asks for your name and its name. These are saved to config.json . AI agents tend to forget and also not personalised to the work you want the agent to do, as well as the agent being in the cloud which brings on the costs and privacy risk. This repo helps you have access to a highly aggressive persoanlised model that does not leave your machine unless you ask it to. Edit config.json to change the model, LoRA settings, or agent behavior. You can also use the CLI: symb config show full config bot token redacted symb config get agent.temperature symb config set agent.temperature 0.7 symb config set telegram.allowed chat ids ' 123456789 ' | Key | Default | Note | |---|---|---| model name | mlx-community/Qwen3-8B-4bit | Base MLX model | assistant name | Symbio | What the assistant calls itself | user name | asked at first run | Your name | agent.max turns | 5 | Max tool rounds per user turn | agent.temperature | 0.1 | Sampling temperature low for deterministic tool use | lora.rank | 8 | LoRA rank adapter width | lora.dropout | 0.1 | LoRA dropout to reduce overfitting | lora.scale | 5.0 | LoRA adapter scale | lora.num layers | 8 | Number of layers to attach adapters to | lora.iters | 50 | LoRA iterations per /train run | lora.max seq length | 2048 | Maximum sequence length during training | lora.learning rate | 1e-4 | LoRA learning rate | lora.save every | 50 | Checkpoint frequency during training | learn.boost factor | 3 | Copies of a correction sample written to training data | learn.batch train iters | 25 | Iterations for threshold-triggered correction training | learn.mistake threshold | 5 | Mistake notes collected before auto-training runs | telegram.bot token | prompted | Telegram bot token from @BotFather | telegram.allowed chat ids | | Chat IDs allowed to use the bot required | After installing pip install -e . the symbio and symb commands are available. During development you can use the symb wrapper script in the repo root. symb Start interactive chat symb chat Same as above symb config Open interactive config editor symb config show Print config.json token redacted symb config get <key Print one value, e.g. agent.temperature symb config set <key <value symb train Run LoRA training symb skill list List saved skills and their adapter status symb skill new <name Create a new skill interactive steps symb skill rm <role Delete a skill, adapter, and training data symb archive Run auto-archive for idle notes/adapters symb archive --dry-run Preview what would be archived symb archive --restore note|adapter <name symb gateway status Check Telegram gateway readiness symb gateway start Start the Telegram bot symb gateway stop Stop a running gateway Legacy python main.py flags still work: python main.py --telegram python main.py --train Run Symbio as a Telegram bot so you can chat from your phone: symb gateway start Legacy equivalent: python main.py --telegram Check gateway readiness first: symb gateway status On first run you will be prompted for a bot token from @BotFather https://t.me/botfather . The token is saved to config.json . For better security, set the environment variable SYMBIO TELEGRAM TOKEN instead; it overrides the config file. You must add your Telegram chat ID to telegram.allowed chat ids : symb config set telegram.allowed chat ids ' 123456789 ' Send any message to the bot, then copy the chat ID from the refusal message if you haven't set it yet. Dangerous actions from Telegram — blocked shell commands, new browser domains, Python code, config changes, cron jobs, digest, and training — ask for approval via an inline keyboard before running. | Command | Description | |---|---| /start | Welcome message | /help | List available commands | /ping | Last turn latency breakdown | /status | Model, adapter, data, and last turn timings | /golden | Run golden-set regression check | /train | Start LoRA training | /selfcheck | Verify enabled features and auto-fix safe issues | /setup | How to change configuration | /tools | Toggle tool groups | /cancel | Clear the current session | | Command | Description | |---|---| /quit | Exit the chat | /save | Save the current conversation to training data | /train | Run LoRA fine-tuning and reload the adapter | /learn | Manually learn from your last correction auto-learn is on by default | /digest | Convert notes into training samples | /note title | Create a markdown note | /notes | List saved notes | /new-skill <name | Create a skill note and start training a worker adapter | /skills | List saved skill notes | /skill-adapters | List skill adapters and their training/idle status | /archive | Archive idle notes/adapters or preview with --dry-run | | /restore note | adapter | /status | Show model, adapter, notes, and session info | /selfcheck | Verify enabled features and auto-fix safe issues | /setup | Re-run the setup wizard names, model, features | /compact | Compress memory/profile store and archive the original | /model | List model presets | /model <preset | Switch to a named model preset restart to load | /run <cmd | Run a sandboxed shell command | /forget last | Remove the last exchange from history | /prune | Remove stale adapter checkpoints | Symbio detects natural corrections automatically and turns them into training data without you typing /learn . Instead of training on every single correction, it saves each mistake as a markdown note in notes/mistakes/ and only fine-tunes once enough notes have accumulated. Typical flow: You: What is my name? Symbio: Your name is Bob. You: No, I'm Alice. Symbio: Your name is Alice. System Correction detected correction phrase . Saved mistake note: 20260715 123456 What is my name.md 1/5 mistake note s collected. Training will run after 4 more correction s . Symbio will: - Detect correction phrases "No, ...", "Actually ...", "That's wrong", etc. or an exact repeat of your last question. - Extract the original question, the wrong answer, the user's correction, and the corrected answer. - Save them as a markdown note in notes/mistakes/ . - When learn.mistake threshold default 5 notes have accumulated, digest them into training data/train.jsonl and run a short LoRA update learn.batch train iters , default 25 . - Archive the used mistake notes to notes/mistakes/archive/ and reload the adapter. The /learn command is still available to force a mistake note from the last correction, but it is no longer required. The same mistake-note pipeline also captures a second, fully automatic pattern that needs no user involvement at all: a tool call that fails, immediately followed by one that works. This is exactly the "wrong command, try the right one" pattern already hand-seeded into every install's base training data — You: Open Chrome. Symbio: <cmd chrome</cmd Tool: run command Observation Command 'chrome' exited error. Output: Command not found: chrome Symbio: 'chrome' isn't a command here — trying the native way. <cmd open -a 'Google Chrome'</cmd Learn Tool mistake captured: 20260721 213045 System observation Command chrome.md — except now it's learned from real usage, not just the seed examples. It feeds into the exact same notes/mistakes/ → threshold → digest → guarded-training pipeline as conversational corrections above, so both count toward the same learn.mistake threshold . Nothing is saved if the model keeps failing without ever finding a working alternative within the turn — only a confirmed fix gets captured. Tune the behaviour in config.json : | Key | Default | Note | |---|---|---| learn.enabled | true | Enable correction learning | learn.auto | true | Detect corrections automatically | learn.auto train | true | Run the fine-tune automatically when the threshold is reached | learn.mistake threshold | 5 | Number of mistake notes before a batch fine-tune runs | learn.batch train iters | 25 | LoRA iterations for the threshold-triggered batch update | learn.boost factor | 3 | Copies of each correction sample written per mistake note | learn.correction phrases | ... | Phrases that trigger correction detection | Symbio uses LoRA Low-Rank Adaptation via Apple's MLX-LM framework. The base model weights stay frozen; only small adapter matrices are trained on curated conversation, notes, and corrections. Training is invoked through the official mlx lm lora CLI: symb train full pass using lora.iters The resulting adapter is saved to adapters/ and loaded automatically on the next start. | Setting | Default | What it controls | |---|---|---| lora.rank | 8 | Width of the low-rank matrices | lora.num layers | 8 | How many transformer layers get adapters | lora.scale | 5.0 | Adapter output scaling | lora.dropout | 0.1 | Dropout for regularization | lora.learning rate | 1e-4 | Training step size | lora.iters | 50 | Iterations for /train | lora.max seq length | 2048 | Training context length | lora.save every | 50 | Checkpoint frequency | Every LoRA update /train , the train adapter tool, the auto-training that follows enough corrections, or the end-of-session prompt is checked against a small, fixed golden set — prompts that exercise behavior baked into every install's seed training data: stating its own name, not confusing itself with the user, emitting the right tool tag for code/notes/reminders/search, and not degenerating into repeated phrases. Each check is single-turn and side-effect-free no tool is actually executed , so it's safe to run automatically. Symbio grades the golden set before training the baseline and again after reloading the new adapter. If a case that passed before now fails, it's a regression , and the previous adapter is restored automatically: Golden Regression: 2 case s newly failing run code for math, web search unknown . Golden Rolled back to the previous adapter. Run it manually anytime with /golden to see the current pass/fail breakdown without training. | Key | Default | Note | |---|---|---| learn.golden set enabled | true | Grade every LoRA update against the golden set | learn.golden rollback on regression | true | Automatically restore the previous adapter on a regression | learn.golden regression threshold | 0 | Newly-failing cases allowed before it counts as a regression | learn.golden max tokens | 150 | Max tokens generated per golden-set case | If a trained adapter exists on disk but the current session isn't using it most commonly after switching model name to something the adapter isn't compatible with , Symbio tracks how long it's sat unused. Past learn.adapter idle days , it asks once whether to remove it: A saved LoRA adapter hasn't been used in 45 day s not loaded with the current model . Remove it to free up space? y/N : Answering yes deletes it; declining or saying "keep" both just leave it alone and reset the grace period, so the reminder won't repeat until it's been idle that long again. Nothing is ever removed without an explicit yes. Check the current status anytime with /status . | Key | Default | Note | |---|---|---| learn.adapter idle reminder enabled | true | Ask about removing an adapter that's gone unused | learn.adapter idle days | 30 | Days unused before the reminder fires | 0731.1.mov One model doing everything — from picking a browser click to answering a factual question — means every micro-decision pays the cost of the headmaster's full system prompt and persona. Symbio can instead hand a bounded sub-task off to a smaller, faster worker model, and each worker can be fine-tuned independently on its own narrow task, with its own adapter, separate from the headmaster's. This is off by default dispatch.enabled: false — it loads and runs additional models on your machine, a bigger resource commitment than anything else here, so it's opt-in. The headmaster requests delegation the same way it requests any other tool: <delegate role='summarize' the full text to condense</delegate or the Hermes form: <tool call {"name": "delegate task", "arguments": {"role": "summarize", "task": "..."}}</tool call . symbio/app/worker models.json is the catalog of available workers — model, role, description, rough memory footprint. Ships with two roles: | Role | What it does | |---|---| summarize | Condenses page/document text handed off by the headmaster | browser | Picks the next click/type/scroll action from the current page text, in a bounded loop, using the same BrowserSession the headmaster's own browser tools drive | Workers load lazily on first use and are evicted LRU-style once dispatch.max resident workers is exceeded, or after sitting idle past dispatch.worker idle unload minutes — sequential by default one resident worker to fit alongside the headmaster on a typical machine, but this is a real, working setting: raise max resident workers if you have the RAM to keep several loaded at once. Every delegated task's input, output pair is recorded as a training sample under that worker's own data directory training data/workers/<role / — real usage builds the corpus. Training a worker reuses the exact golden-set-guarded-rollback machinery the headmaster's own /train uses: a small, role-scoped golden set e.g. "does the browser worker still reply with a known action verb" is checked before and after training, and a regression rolls the worker's adapter back automatically, the same way guarded train protects the headmaster's. Worker adapters live under adapters/workers/<role / , fully separate from the headmaster's own adapters/ . | Key | Default | Note | |---|---|---| dispatch.enabled | false | Turn on delegation | dispatch.max resident workers | 1 | How many worker models can be loaded at once | dispatch.worker idle unload minutes | 10 | Unload a worker after this long unused | dispatch.max worker rounds | 4 | Round cap for a multi-step worker task e.g. browser | dispatch.worker golden set enabled | true | Golden-check a worker's adapter around training | dispatch.worker golden rollback on regression | true | Auto-rollback a worker's adapter on regression | Enabling dispatch also needs "delegate" in tools.enabled groups — it's included by default going forward, but an existing config.json written before this feature won't have picked it up automatically; add it with /config set tools.enabled groups ' ... ' if delegation seems to silently do nothing. "My name is Alice." "Call me Bob." "You can call me Charlie." "From now on call me Dana." "Change my name to Eve." "I go by Frank." "Call yourself Jarvis." "I will call you Friday." "I'm going to call you HAL." "Change your name to Jeeves." "Set your name as Alfred." Note: "Your name is X"is intentionallynottreated as an assistant rename because small models often confuse it with the user's name. Symbio understands two ways to call tools: - Legacy XML tags : <note title="User Preference" The user likes coffee.</note — save a note <cmd ls</cmd — run a sandboxed command legacy, still supported <digest / / <train / — digest notes or train - Hermes JSON-in-XML preferred : <tool call {"name": "read file", "arguments": {"path": "config.json"}}</tool call <tool call {"name": "terminal", "arguments": {"cmd": "ls -la"}}</tool call <tool call {"name": "note", "arguments": {"action": "add", "target": "note", "content": "The user likes coffee."}}</tool call terminal and execute code are best-effort sandboxes. They run with the privileges of the user who started the program and are scoped to the project directory. execute code requires the script to import from symbio tools or the backward-compatible caine tools alias and blocks known dangerous imports.- Do not paste untrusted code into the agent without reviewing it first. - And also do pay attention to the 'Do you wanna yes/no' questions those are there to keep you from having Symbio do random stuff without your consent because you might not want to do it. The project is organized as a symbio/ Python package with a thin main.py wrapper: . ├── main.py Delegates to the modern CLI in symbio/app/cli.py ├── symbio/ │ ├── constants.py Paths, DEFAULT CONFIG │ ├── app/ │ │ ├── cli.py symbio / symb command-line interface │ │ ├── chat.py ChatSession, agent loop, slash commands │ │ ├── config.py Defaults, loading, redaction, token prompt │ │ ├── training.py Training data and LoRA fine-tuning via mlx lm │ │ ├── learn.py Correction detection and batch learning │ │ ├── golden.py Golden set: regression checks around every LoRA update │ │ ├── dispatch.py MoA: WorkerPool, delegated tasks, worker fine-tuning │ │ ├── worker models.json Catalog of available worker models/roles │ │ ├── memory.py Notes, memory, profile management │ │ ├── sandbox.py Sandboxed commands and Python execution │ │ ├── computer.py Browser automation helpers │ │ ├── cron.py Scheduled jobs and reminders │ │ ├── telegram.py Telegram bot gateway │ │ ├── tooling.py Tag parsing and tool stripping | | ├── prompts.py just the prompt idk nothing special | | └── skills.py saves the skills as adapters as well as manage the notes/ │ └── utils.py Shared helpers ├── rag.py Lightweight keyword-based RAG ├── README.md ├── docs/ │ └── adapter-marketplace.md Design doc, not yet implemented ├── config.json User configuration ├── models.json Model presets ├── notes/ Markdown notes / memory ├── training data/ train.jsonl and valid.jsonl workers/<role / for MoA workers ├── adapters/ LoRA adapter weights workers/<role / for MoA workers ├── logs/ Session logs ├── sessions/ Session stores ├── screenshots/ Browser screenshots └── sandbox/ Scratch space for code execution We are actively looking for help on: CUDA port — MLX is Apple Silicon only. A PyTorch or Transformers backend would let Symbio run on NVIDIA/AMD hardware. llama.cpp backend — Support GGUF models through llama.cpp for broader model coverage and lower memory use. LoRA optimization — Faster adapter swaps, gradient checkpointing, and memory-efficient training. Refactoring — Cleaner separation between inference, tools, training, and storage; better test coverage. Sparse / quantized adapters — Experiment with QLoRA, 8-bit/4-bit base models, and sparse LoRA updates. See CONTRIBUTING.md /huyedits/Symbio/blob/main/CONTRIBUTING.md for setup, testing, and how to open issues/PRs. future projection < - Add MCP Model Context Protocol - Add More Tools — live browser <browse / <click / <type / <scroll , <skill , permission-gated sandbox - Self correction when hallucinating - Be able to learn new skills on the fly - Remember new info found from web research — auto-saved as Learned: notes, trained in on digest - Add Telegram bot — full tool loop with inline-keyboard approval for dangerous actions - Mixture of agents — headmaster delegates bounded sub-tasks to smaller, independently fine-tunable worker models dispatch.enabled , off by default - Adapter marketplace — design doc: docs/adapter-marketplace.md /huyedits/Symbio/blob/main/docs/adapter-marketplace.md ; not yet implemented - Add Other Messaging Platforms - Prune Old Weights Future Milestone Apache 2.0