automindx — IAML

Intelligent Autonomous Machine Learning · "I Am Machine Learning"

Professor Codephreak — an expert in machine learning, computer science, and secure, modular programming — as a local, persona-driven language-model environment. Project codename: codephreak.

---

⚡ Quickstart — the modern path (Ollama)

The classic uiux.py / hfUIUX.py load a multi-GB model before the UI renders, so on modest hardware the window may never appear. ollama_codephreak.py inverts that: the model runs in the Ollama daemon and the Gradio UI loads instantly (degrading gracefully when no model is present). It ships a model picker (local + :cloud), live streaming, and a token counter (prompt + completion + tok/s).

pip install -r requirements-ollama.txt   # gradio + requests only — no torch
ollama serve                             # in another terminal
ollama pull qwen3:0.6b                    # small local model for modest hardware
python3 ollama_codephreak.py             # → http://localhost:7860

codephreak is AGLM — an augmentation layer, not a model. The full AGLM cognitive console (Socratic, Logic, BDI, MASTERMIND, …) and an AI-SDK streaming participant UI live at GATERAGE/aglm.

📚 docs/NAV.md — full documentation index  ·  📖 TECHNICAL.md — module map & entrypoints  ·  🔧 AUDIT.md — audit findings & fixes

---

Entrypoints

Path	File	Loads	Requires
AI SDK console (definitive)	codephreak-console/ + codephreak.py	instantly	Ollama + Node (see CODEPHREAK_CONSOLE.md)
Modern Gradio	ollama_codephreak.py	instantly	a running Ollama daemon
Legacy GGML (CPU)	hfapp.py	after model download	llama-cpp-python + ~4 GB model
Legacy transformers	uiux.py / hfUIUX.py	after model load	torch + transformers

The Professor Codephreak AI SDK console is the flagship UI: streaming chat on the Vercel AI SDK v7 with gpt-oss by default, reasoning, a live token / tok-s counter, advanced and scientific sampling, a .persona creator (Codephreak, automindX, jAImla, Savante, …), .history, export/copy, ❤/💔 feedback into the self-improving codephreak.py, and a living WebGL substrate that fluxes while thinking. codephreak also has read-only filesystem access (tool-calling: list_files/read_file/grep) so it answers about the actual code, and decomposes large tasks into sub-tasks completed one at a time.

./scripts/run_codephreak_console.sh      # engine :5001 + console :3100

---

Using automindX

1. The console (recommended)

Prerequisites: Ollama and Node.js 18+.

ollama serve                     # start the model daemon (another terminal)
ollama pull qwen3:0.6b           # a small local model for testing (optional)
./scripts/run_codephreak_console.sh

Open http://localhost:3100. Then:

• Chat — type a message and press Enter. Responses stream live, with a collapsible reasoning trace, a live token / tok-s counter, and a substrate that gently pulses while codephreak thinks.
• Pick a model (top bar) — local models, or free :cloud models (gpt-oss:120b-cloud is the default). Gated models show a sign-in link; Autonomous mode auto-pulls a model you select but haven't installed.
• Ask about the actual code — codephreak can read the real project files, so “list all project files”, “read services/model_service.py”, or “audit the console for security issues” return grounded answers (it shows the tool calls in the reasoning stream). Big tasks are auto-split into sub-tasks.
• .persona — switch or edit personas (Professor Codephreak, automindX, jAImla, Savante, Sentinel, Architect, Mentor) or create your own; the edited prompt is what's sent.
• Advanced / Scientific — tune sampling (temperature, top-p/k, mirostat, num_ctx, …). Preferences — avatar, name, accent colour. .history — reopen / export / delete past chats.
• ❤ / 👍 / 👎 / 💔 — rate replies; codephreak.py learns directives from your feedback and folds them into the persona over time.
• sAGI (Preferences → toggle) — opens a tab where a self-building sAGI grows one module at a time into the sagi/ package.

2. From Python (the service layer)

from automind import chat
print(chat("Write a production-ready Python rate limiter."))   # one-shot
sid = None
out = chat("Remember my name is Ada.", full=True); sid = out["session_id"]
print(chat("What's my name?", session_id=sid))                 # continues with memory

Memory persists to SQLite by default (set AUTOMINDX_MEMORY_BACKEND=pgvector for RAGE semantic memory). See docs/SERVICES.md.

3. Command line

python3 -m services.inference_orchestrator      # interactive REPL
python3 -m services.self_audit                  # codephreak audits the real code
python3 -m services.self_audit --file services/model_service.py
python3 codephreak.py                           # self-improving engine (:5001, /healthz)
python3 sagi_build.py --steps 3                 # headless sAGI self-build

4. Gradio (no Node)

pip install -r requirements-ollama.txt && python3 ollama_codephreak.py   # → :7860

Config is env-driven (OLLAMA_HOST, AUTOMINDX_MODEL, CODEPHREAK_MODEL, AUTOMINDX_MEMORY_BACKEND, …) — see docs/SERVICES.md.

Modules

Documentation: codephreak = uiux.py + memory.py + automind.py + aglm.py

Module	Responsibility
services/	The decoupled service layer (see SERVICES.md): MemoryService (SQLite) · RageMemory (pgvector/pgvectorscale) · get_memory() factory · ModelService (lazy Ollama client) · InferenceOrchestrator (sanitize → RAG recall → infer → persist → JSON logs) · ModelRegistry (versioned config + digest integrity) · secrets (keyring→env) · self_audit (codephreak reads the real source).
automind.py	Thin façade — chat(user_input, session_id) delegates to the service layer; also exports the codephreak persona + format_to_llama_chat_style.
codephreak.py	Self-improving persona engine — learns directives from realtime 👍/👎 feedback (GET /healthz, POST /feedback).
memory.py	Legacy JSON conversation memory (./memory/*.json); the service layer uses SQLite/pgvector.
llama_model.py	LlamaModel — loads a local dir or a Hugging Face id (device_map="auto"). (was aglm.py)
ollama_codephreak.py	Modern Ollama-backed Gradio chat with model picker + token counter.
aglm/	The Autonomous General Learning Model package (PODA cycle · beliefs · autonomous loop), migrated from GATERAGE/aglm. See AGLM.md.
automind_aglm.py	Wires aGLM to codephreak's feedback so automindX self-refines its persona.

Reality note. automindX runs models in the Ollama daemon (local + free :cloud), not an in-process torch checkpoint. Self-improvement is prompt-space (feedback → learned directives), not weight-space training. The service layer, registry, secrets, and self-audit are real and tested (python3 -m pytest).

---

Install

The scripts/automindx.install script sets up the full environment (tested on Ubuntu 22.04, Linux Mint 21.2, and Mandrake Linux):

chmod +x scripts/automindx.install && ./scripts/automindx.install

Legacy GGML run

Auto-downloads llama2-7b-chat-codeCherryPop-qLoRA-GGML and launches Professor Codephreak with an agenda to build the automindx deployment environment.

python3 uiux.py \
  --model_name="TheBloke/llama2-7b-chat-codeCherryPop-qLoRA-GGML" \
  --model_type="ggml" \
  --file_name="llama-2-7b-chat-codeCherryPop.ggmlv3.q4_1.bin" \
  --save_history

Note. The original model has a hard 4096-token context. The token-aware sliding window in context4096.py (wired into hfUIUX.py) fixes the old overflow crash; modern models via Ollama sidestep it entirely with a larger num_ctx. See CONTEXT.md.

---

Links

• Professor Codephreak — the origin and the evolving author; also runs the team of OpenAI agents at gpt.pythai.net

• github.com/GATERAGE — the organization

  • GATERAGE/aglm — Autonomous General Learning Model (package + consoles)
  • GATERAGE/RAGE — retrieval substrate (memory); sAGI saves its memory trees here (pgvectorscale · rage.pythai.net)
  • GATERAGE/mastermind — strategic orchestrator; mastermind.pythai.net is the warcouncil interaction to the boardroom

• github.com/aiterm — augmented intelligence terminal, the terminal-AI arm of the PYTHAI / Professor Codephreak ecosystem (pythai.net): a curated hub of terminal-first AI/CLI tooling on the thesis that the terminal is the primary interface for AI-augmented development. automindX embodies this directly — the console's ⌘ Terminal lets you call Claude from your subscription CLI and work on the dapp from inside the dapp.

• Professor-Codephreak/gitmind — the git-like memory that sAGI's in-runtime gitmind ports (local/global commit tiers; trees embedded into RAGE — see docs/savethetrees.md)

• Original codephreak (historical) — https://github.com/Professor-Codephreak/automind

• mindX — the production system aGLM was distilled from · docs: mindx.pythai.net/docs.html

• rage.pythai.net — the public broadcast layer ("public noise"), where mindX publishes in the voice of mindX via its AuthorAgent; Professor Codephreak has published there occasionally

• Songs — the personas appear in music by Mag Magnus (authored by Magnusson):

     

RAGE remembers · aGLM decides · MASTERMIND orchestrates.

This repo is under active development — Professor Codephreak keeps evolving.